aid
stringlengths 9
15
| mid
stringlengths 7
10
| abstract
stringlengths 78
2.56k
| related_work
stringlengths 92
1.77k
| ref_abstract
dict |
|---|---|---|---|---|
1907.08302 | 2964151964 | With the demand to process ever-growing data volumes, a variety of new data stream processing frameworks have been developed. Moving an implementation from one such system to another, e.g., for performance reasons, requires adapting existing applications to new interfaces. Apache Beam addresses these high substitution costs by providing an abstraction layer that enables executing programs on any of the supported streaming frameworks. In this paper, we present a novel benchmark architecture for comparing the performance impact of using Apache Beam on three streaming frameworks: Apache Spark Streaming, Apache Flink, and Apache Apex. We find significant performance penalties when using Apache Beam for application development in the surveyed systems. Overall, usage of Apache Beam for the examined streaming applications caused a high variance of query execution times with a slowdown of up to a factor of 58 compared to queries developed without the abstraction layer. All developed benchmark artifacts are publicly available to ensure reproducible results. | The mentioned data sender emits data to the DSPS, which is mostly car position reports. Depending on the overall situation on the expressways, car position reports may require the DSPS to create an output or not. Next to car position reports, the remaining input data represent an explicit query which always requires an answer. Linear Road defines four distinct queries, whereas the query lastly presented in @cite_42 was skipped in the two presented implementations due to complexity reasons. | {
"cite_N": [
"@cite_42"
],
"mid": [
"2112215401"
],
"abstract": [
"This paper specifies the Linear Road Benchmark for Stream Data Management Systems (SDMS). Stream Data Management Systems process streaming data by executing continuous and historical queries while producing query results in real-time. This benchmark makes it possible to compare the performance characteristics of SDMS' relative to each other and to alternative (e.g., Relational Database) systems. Linear Road has been endorsed as an SDMS benchmark by the developers of both the Aurora [1] (out of Brandeis University, Brown University and MIT) and STREAM [8] (out of Stanford University) stream systems. Linear Road simulates a toll system for the motor vehicle expressways of a large metropolitan area. The tolling system uses \"variable tolling\" [6, 11, 9]: an increasingly prevalent tolling technique that uses such dynamic factors as traffic congestion and accident proximity to calculate toll charges. Linear Road specifies a variable tolling system for a fictional urban area including such features as accident detection and alerts, traffic congestion measurements, toll calculations and historical queries. After specifying the benchmark, we describe experimental results involving two implementations: one using a commercially available Relational Database and the other using Aurora. Our results show that a dedicated Stream Data Management System can outperform a Relational Database by at least a factor of 5 on streaming data applications."
]
} |
1907.08302 | 2964151964 | With the demand to process ever-growing data volumes, a variety of new data stream processing frameworks have been developed. Moving an implementation from one such system to another, e.g., for performance reasons, requires adapting existing applications to new interfaces. Apache Beam addresses these high substitution costs by providing an abstraction layer that enables executing programs on any of the supported streaming frameworks. In this paper, we present a novel benchmark architecture for comparing the performance impact of using Apache Beam on three streaming frameworks: Apache Spark Streaming, Apache Flink, and Apache Apex. We find significant performance penalties when using Apache Beam for application development in the surveyed systems. Overall, usage of Apache Beam for the examined streaming applications caused a high variance of query execution times with a slowdown of up to a factor of 58 compared to queries developed without the abstraction layer. All developed benchmark artifacts are publicly available to ensure reproducible results. | The benchmark result for a system is summarized as a so called L-rating. This metric defined by Linear Road expresses how many expressways the system could handle while meeting the defined response time requirements for each query. A higher number of expressways corresponds to a higher data input rate for the SUT. When generating data, the amount of expressways can be configured.s The Linear Road benchmark was applied to the DSPS Aurora @cite_24 and a commercial relational database. Results are presented in the paper. | {
"cite_N": [
"@cite_24"
],
"mid": [
"2112215401"
],
"abstract": [
"This paper specifies the Linear Road Benchmark for Stream Data Management Systems (SDMS). Stream Data Management Systems process streaming data by executing continuous and historical queries while producing query results in real-time. This benchmark makes it possible to compare the performance characteristics of SDMS' relative to each other and to alternative (e.g., Relational Database) systems. Linear Road has been endorsed as an SDMS benchmark by the developers of both the Aurora [1] (out of Brandeis University, Brown University and MIT) and STREAM [8] (out of Stanford University) stream systems. Linear Road simulates a toll system for the motor vehicle expressways of a large metropolitan area. The tolling system uses \"variable tolling\" [6, 11, 9]: an increasingly prevalent tolling technique that uses such dynamic factors as traffic congestion and accident proximity to calculate toll charges. Linear Road specifies a variable tolling system for a fictional urban area including such features as accident detection and alerts, traffic congestion measurements, toll calculations and historical queries. After specifying the benchmark, we describe experimental results involving two implementations: one using a commercially available Relational Database and the other using Aurora. Our results show that a dedicated Stream Data Management System can outperform a Relational Database by at least a factor of 5 on streaming data applications."
]
} |
1907.08302 | 2964151964 | With the demand to process ever-growing data volumes, a variety of new data stream processing frameworks have been developed. Moving an implementation from one such system to another, e.g., for performance reasons, requires adapting existing applications to new interfaces. Apache Beam addresses these high substitution costs by providing an abstraction layer that enables executing programs on any of the supported streaming frameworks. In this paper, we present a novel benchmark architecture for comparing the performance impact of using Apache Beam on three streaming frameworks: Apache Spark Streaming, Apache Flink, and Apache Apex. We find significant performance penalties when using Apache Beam for application development in the surveyed systems. Overall, usage of Apache Beam for the examined streaming applications caused a high variance of query execution times with a slowdown of up to a factor of 58 compared to queries developed without the abstraction layer. All developed benchmark artifacts are publicly available to ensure reproducible results. | The work presented in @cite_11 compares Apache Flink and Apache Spark. The conducted measurements include different queries, a grep query being one of them. One focus area that is analyzed is the scaling behavior with regard to different numbers of nodes in the cluster. However, studying both systems from a data stream processing point of view is out of scope in the performed measurements. | {
"cite_N": [
"@cite_11"
],
"mid": [
"2498111289"
],
"abstract": [
"Big Data analytics has recently gained increasing popularity as a tool to process large amounts of data on-demand. Spark and Flink are two Apache-hosted data analytics frameworks that facilitate the development of multi-step data pipelines using directly acyclic graph patterns. Making the most out of these frameworks is challenging because efficient executions strongly rely on complex parameter configurations and on an in-depth understanding of the underlyingarchitectural choices. Although extensive research has been devoted to improving and evaluating the performance of such analytics frameworks, most of them benchmarkthe platforms against Hadoop, as a baseline, a rather unfair comparison consideringthe fundamentally different design principles. This paper aims to bring some justice in this respect, by directly evaluating the performance of Sparkand Flink. Our goal is to identify and explain the impact of the different architecturalchoices and the parameter configurations on the perceived end-to-end performance. To this end, we develop a methodology for correlating the parameter settings and the operators execution plan with the resource usage. We use this methodologyto dissect the performance of Spark and Flink with several representative batchand iterative workloads on up to 100 nodes. Our key finding is that there none of the two framework outperforms the other for all data types, sizes and job patterns. This paper performs a fine characterization of the cases when each framework is superior, and we highlight how this performance correlates to operators, to resource usage and to the specifics of the internal framework design."
]
} |
1907.08302 | 2964151964 | With the demand to process ever-growing data volumes, a variety of new data stream processing frameworks have been developed. Moving an implementation from one such system to another, e.g., for performance reasons, requires adapting existing applications to new interfaces. Apache Beam addresses these high substitution costs by providing an abstraction layer that enables executing programs on any of the supported streaming frameworks. In this paper, we present a novel benchmark architecture for comparing the performance impact of using Apache Beam on three streaming frameworks: Apache Spark Streaming, Apache Flink, and Apache Apex. We find significant performance penalties when using Apache Beam for application development in the surveyed systems. Overall, usage of Apache Beam for the examined streaming applications caused a high variance of query execution times with a slowdown of up to a factor of 58 compared to queries developed without the abstraction layer. All developed benchmark artifacts are publicly available to ensure reproducible results. | @cite_20 compare Apache Storm, Apache Flink, and Apache Spark Streaming in their paper. Besides describing the architecture of these three systems, the performance is studied in a network traffic analysis scenario. Additionally, the behavior in case of a node failure is investigated. | {
"cite_N": [
"@cite_20"
],
"mid": [
"2498111289"
],
"abstract": [
"Big Data analytics has recently gained increasing popularity as a tool to process large amounts of data on-demand. Spark and Flink are two Apache-hosted data analytics frameworks that facilitate the development of multi-step data pipelines using directly acyclic graph patterns. Making the most out of these frameworks is challenging because efficient executions strongly rely on complex parameter configurations and on an in-depth understanding of the underlyingarchitectural choices. Although extensive research has been devoted to improving and evaluating the performance of such analytics frameworks, most of them benchmarkthe platforms against Hadoop, as a baseline, a rather unfair comparison consideringthe fundamentally different design principles. This paper aims to bring some justice in this respect, by directly evaluating the performance of Sparkand Flink. Our goal is to identify and explain the impact of the different architecturalchoices and the parameter configurations on the perceived end-to-end performance. To this end, we develop a methodology for correlating the parameter settings and the operators execution plan with the resource usage. We use this methodologyto dissect the performance of Spark and Flink with several representative batchand iterative workloads on up to 100 nodes. Our key finding is that there none of the two framework outperforms the other for all data types, sizes and job patterns. This paper performs a fine characterization of the cases when each framework is superior, and we highlight how this performance correlates to operators, to resource usage and to the specifics of the internal framework design."
]
} |
1902.00126 | 2915013032 | We propose a stochastic gradient framework for solving stochastic composite convex optimization problems with (possibly) infinite number of linear inclusion constraints that need to be satisfied almost surely. We use smoothing and homotopy techniques to handle constraints without the need for matrix-valued projections. We show for our stochastic gradient algorithm @math convergence rate for general convex objectives and @math convergence rate for restricted strongly convex objectives. These rates are known to be optimal up to logarithmic factors, even without constraints. We demonstrate the performance of our algorithm with numerical experiments on basis pursuit, a hard margin support vector machines and a portfolio optimization and show that our algorithm achieves state-of-the-art practical performance. | The most prominent work for stochastic optimization problems is stochastic gradient descent (SGD) @cite_17 @cite_13 @cite_5 . Even though SGD is very well studied, it only applies when there does not exist any constraints in the problem template . For the case of simple constraints, @math @math in and almost sure constraints are not present, projected SGD can be used @cite_17 . However, it requires @math to be a projectable set, which does not apply to the general template of which would involve the almost sure constraints in the definition of @math . In the case where @math in is a nonsmooth proximable function @cite_22 studied the convergence of stochastic proximal gradient (SPG) method which utilizes stochastic gradients of @math in addition to the proximal operator of @math . This method generalize projected SGD, however, they cannot handle infinitely many constraints that we consider in since it is not possible to project onto their intersection in general. | {
"cite_N": [
"@cite_5",
"@cite_13",
"@cite_22",
"@cite_17"
],
"mid": [
"2962795635",
"2951196414",
"2500690480",
"793269399"
],
"abstract": [
"Momentum based stochastic gradient methods such as heavy ball (HB) and Nesterov's accelerated gradient descent (NAG) method are widely used in practice for training deep networks and other supervised learning models, as they often provide significant improvements over stochastic gradient descent (SGD). In general, “fast gradient” methods have provable improvements over gradient descent only for the deterministic case, where the gradients are exact. In the stochastic case, the popular explanations for their wide applicability is that when these fast gradient methods are applied in the stochastic case, they partially mimic their exact gradient counterparts, resulting in some practical gain. This work provides a counterpoint to this belief by proving that there are simple problem instances where these methods cannot outperform SGD despite the best setting of its parameters. These negative problem instances are, in an informal sense, generic; they do not look like carefully constructed pathological instances. These results suggest (along with empirical evidence) that HB or NAG's practical performance gains are a by-product of minibatching. Furthermore, this work provides a viable (and provable) alternative, which, on the same set of problem instances, significantly improves over HB, NAG, and SGD's performance. This algorithm, denoted as ASGD, is a simple to implement stochastic algorithm, based on a relatively less popular version of Nesterov's AGD. Extensive empirical results in this paper show that ASGD has performance gains over HB, NAG, and SGD.",
"Stochastic gradient descent (SGD) is a simple and popular method to solve stochastic optimization problems which arise in machine learning. For strongly convex problems, its convergence rate was known to be O( (T) T), by running SGD for T iterations and returning the average point. However, recent results showed that using a different algorithm, one can get an optimal O(1 T) rate. This might lead one to believe that standard SGD is suboptimal, and maybe should even be replaced as a method of choice. In this paper, we investigate the optimality of SGD in a stochastic setting. We show that for smooth problems, the algorithm attains the optimal O(1 T) rate. However, for non-smooth problems, the convergence rate with averaging might really be ( (T) T), and this is not just an artifact of the analysis. On the flip side, we show that a simple modification of the averaging step suffices to recover the O(1 T) rate, and no other change of the algorithm is necessary. We also present experimental results which support our findings, and point out open problems.",
"We consider stochastic strongly convex optimization with a complex inequality constraint. This complex inequality constraint may lead to computationally expensive projections in algorithmic iterations of the stochastic gradient descent (SGD) methods. To reduce the computation costs pertaining to the projections, we propose an Epoch-Projection Stochastic Gradient Descent (Epro-SGD) method. The proposed Epro-SGD method consists of a sequence of epochs; it applies SGD to an augmented objective function at each iteration within the epoch, and then performs a projection at the end of each epoch. Given a strongly convex optimization and for a total number of @math iterations, Epro-SGD requires only @math projections, and meanwhile attains an optimal convergence rate of @math , both in expectation and with a high probability. To exploit the structure of the optimization problem, we propose a proximal variant of Epro-SGD, namely Epro-ORDA, based on the optimal regularized dual averaging method. We apply the proposed methods on real-world applications; the empirical results demonstrate the effectiveness of our methods.",
"Stochastic gradient descent (SGD) holds as a classical method to build large scale machine learning models over big data. A stochastic gradient is typically calculated from a limited number of samples (known as mini-batch), so it potentially incurs a high variance and causes the estimated parameters bounce around the optimal solution. To improve the stability of stochastic gradient, recent years have witnessed the proposal of several semi-stochastic gradient descent algorithms, which distinguish themselves from standard SGD by incorporating global information into gradient computation. In this paper we contribute a novel stratified semi-stochastic gradient descent (S3GD) algorithm to this nascent research area, accelerating the optimization of a large family of composite convex functions. Though theoretically converging faster, prior semi-stochastic algorithms are found to suffer from high iteration complexity, which makes them even slower than SGD in practice on many datasets. In our proposed S3GD, the semi-stochastic gradient is calculated based on efficient manifold propagation, which can be numerically accomplished by sparse matrix multiplications. This way S3GD is able to generate a highly-accurate estimate of the exact gradient from each mini-batch with largely-reduced computational complexity. Theoretic analysis reveals that the proposed S3GD elegantly balances the geometric algorithmic convergence rate against the space and time complexities during the optimization. The efficacy of S3GD is also experimentally corroborated on several large-scale benchmark datasets."
]
} |
1902.00126 | 2915013032 | We propose a stochastic gradient framework for solving stochastic composite convex optimization problems with (possibly) infinite number of linear inclusion constraints that need to be satisfied almost surely. We use smoothing and homotopy techniques to handle constraints without the need for matrix-valued projections. We show for our stochastic gradient algorithm @math convergence rate for general convex objectives and @math convergence rate for restricted strongly convex objectives. These rates are known to be optimal up to logarithmic factors, even without constraints. We demonstrate the performance of our algorithm with numerical experiments on basis pursuit, a hard margin support vector machines and a portfolio optimization and show that our algorithm achieves state-of-the-art practical performance. | A line of work that is known as alternating projections, focus on applying random projections for solving problems that are involving the intersection of infinite number of sets. In particular, these methods focus on the following template Here, the feasible set @math consists of the intersection of a possibly infinite number of convex sets. The case when @math which corresponds to the convex feasibility problem is studied in @cite_19 . For this particular setting, the authors combine the smoothing technique with minibatch SGD, leading to a stochastic alternating projection algorithm having linear convergence. | {
"cite_N": [
"@cite_19"
],
"mid": [
"2783403261"
],
"abstract": [
"Finding a point in the intersection of a collection of closed convex sets, that is the convex feasibility problem, represents the main modeling strategy for many computational problems. In this paper we analyze new stochastic reformulations of the convex feasibility problem in order to facilitate the development of new algorithmic schemes. We also analyze the conditioning problem parameters using certain (linear) regularity assumptions on the individual convex sets. Then, we introduce a general random projection algorithmic framework, which extends to the random settings many existing projection schemes, designed for the general convex feasibility problem. Our general random projection algorithm allows to project simultaneously on several sets, thus providing great flexibility in matching the implementation of the algorithm on the parallel architecture at hand. Based on the conditioning parameters, besides the asymptotic convergence results, we also derive explicit sublinear and linear convergence rates for this general algorithmic framework."
]
} |
1902.00126 | 2915013032 | We propose a stochastic gradient framework for solving stochastic composite convex optimization problems with (possibly) infinite number of linear inclusion constraints that need to be satisfied almost surely. We use smoothing and homotopy techniques to handle constraints without the need for matrix-valued projections. We show for our stochastic gradient algorithm @math convergence rate for general convex objectives and @math convergence rate for restricted strongly convex objectives. These rates are known to be optimal up to logarithmic factors, even without constraints. We demonstrate the performance of our algorithm with numerical experiments on basis pursuit, a hard margin support vector machines and a portfolio optimization and show that our algorithm achieves state-of-the-art practical performance. | Stochastic forward-backward algorithms can also be applied to solve . However, the papers introducing those very general algorithms focused on proving convergence and did not present convergence rates @cite_0 @cite_29 @cite_20 . There are some other works that focus on @cite_2 @cite_10 @cite_25 where the authors assume the number of constraints is finite, which is more restricted than our setting. | {
"cite_N": [
"@cite_29",
"@cite_0",
"@cite_2",
"@cite_10",
"@cite_25",
"@cite_20"
],
"mid": [
"2963692017",
"1837565340",
"2259139337",
"1967138577",
"1987083649",
"2952274722"
],
"abstract": [
"We investigate the asymptotic behavior of a stochastic version of the forward-backward splitting algorithm for finding a zero of the sum of a maximally monotone set-valued operator and a cocoercive operator in Hilbert spaces. Our general setting features stochastic approximations of the cocoercive operator and stochastic perturbations in the evaluation of the resolvents of the set-valued operator. In addition, relaxations and not necessarily vanishing proximal parameters are allowed. Weak and strong almost sure convergence properties of the iterates is established under mild conditions on the underlying stochastic processes. Leveraging these results, we also establish the almost sure convergence of the iterates of a stochastic variant of a primal-dual proximal splitting method for composite minimization problems.",
"The forward-backward algorithm is a powerful tool for solving optimization problems with an additively separable and smooth plus nonsmooth structure. In the convex setting, a simple but ingenious acceleration scheme developed by Nesterov improves the theoretical rate of convergence for the function values from the standard @math down to @math . In this short paper, we prove that the rate of convergence of a slight variant of Nesterov's accelerated forward-backward method, which produces convergent sequences, is actually @math , rather than @math . Our arguments rely on the connection between this algorithm and a second-order differential inclusion with vanishing damping.",
"A number of recent works have emphasized the prominent role played by the Kurdyka-źojasiewicz inequality for proving the convergence of iterative algorithms solving possibly nonsmooth nonconvex optimization problems. In this work, we consider the minimization of an objective function satisfying this property, which is a sum of two terms: (i) a differentiable, but not necessarily convex, function and (ii) a function that is not necessarily convex, nor necessarily differentiable. The latter function is expressed as a separable sum of functions of blocks of variables. Such an optimization problem can be addressed with the Forward---Backward algorithm which can be accelerated thanks to the use of variable metrics derived from the Majorize---Minimize principle. We propose to combine the latter acceleration technique with an alternating minimization strategy which relies upon a flexible update rule. We give conditions under which the sequence generated by the resulting Block Coordinate Variable Metric Forward---Backward algorithm converges to a critical point of the objective function. An application example to a nonconvex phase retrieval problem encountered in signal image processing shows the efficiency of the proposed optimization method.",
"In view of the minimization of a nonsmooth nonconvex function f, we prove an abstract convergence result for descent methods satisfying a sufficient-decrease assumption, and allowing a relative error tolerance. Our result guarantees the convergence of bounded sequences, under the assumption that the function f satisfies the Kurdyka–Łojasiewicz inequality. This assumption allows to cover a wide range of problems, including nonsmooth semi-algebraic (or more generally tame) minimization. The specialization of our result to different kinds of structured problems provides several new convergence results for inexact versions of the gradient method, the proximal method, the forward–backward splitting algorithm, the gradient projection and some proximal regularization of the Gauss–Seidel method in a nonconvex setting. Our results are illustrated through feasibility problems, or iterative thresholding procedures for compressive sensing.",
"In this paper, we generalize the well-known Nesterov's accelerated gradient (AG) method, originally designed for convex smooth optimization, to solve nonconvex and possibly stochastic optimization problems. We demonstrate that by properly specifying the stepsize policy, the AG method exhibits the best known rate of convergence for solving general nonconvex smooth optimization problems by using first-order information, similarly to the gradient descent method. We then consider an important class of composite optimization problems and show that the AG method can solve them uniformly, i.e., by using the same aggressive stepsize policy as in the convex case, even if the problem turns out to be nonconvex. We demonstrate that the AG method exhibits an optimal rate of convergence if the composite problem is convex, and improves the best known rate of convergence if the problem is nonconvex. Based on the AG method, we also present new nonconvex stochastic approximation methods and show that they can improve a few existing rates of convergence for nonconvex stochastic optimization. To the best of our knowledge, this is the first time that the convergence of the AG method has been established for solving nonconvex nonlinear programming in the literature.",
"In this paper, we generalize the well-known Nesterov's accelerated gradient (AG) method, originally designed for convex smooth optimization, to solve nonconvex and possibly stochastic optimization problems. We demonstrate that by properly specifying the stepsize policy, the AG method exhibits the best known rate of convergence for solving general nonconvex smooth optimization problems by using first-order information, similarly to the gradient descent method. We then consider an important class of composite optimization problems and show that the AG method can solve them uniformly, i.e., by using the same aggressive stepsize policy as in the convex case, even if the problem turns out to be nonconvex. We demonstrate that the AG method exhibits an optimal rate of convergence if the composite problem is convex, and improves the best known rate of convergence if the problem is nonconvex. Based on the AG method, we also present new nonconvex stochastic approximation methods and show that they can improve a few existing rates of convergence for nonconvex stochastic optimization. To the best of our knowledge, this is the first time that the convergence of the AG method has been established for solving nonconvex nonlinear programming in the literature."
]
} |
1902.00126 | 2915013032 | We propose a stochastic gradient framework for solving stochastic composite convex optimization problems with (possibly) infinite number of linear inclusion constraints that need to be satisfied almost surely. We use smoothing and homotopy techniques to handle constraints without the need for matrix-valued projections. We show for our stochastic gradient algorithm @math convergence rate for general convex objectives and @math convergence rate for restricted strongly convex objectives. These rates are known to be optimal up to logarithmic factors, even without constraints. We demonstrate the performance of our algorithm with numerical experiments on basis pursuit, a hard margin support vector machines and a portfolio optimization and show that our algorithm achieves state-of-the-art practical performance. | Another related work is @cite_15 where the authors apply Nesterov's smoothing to . However, this work does not apply to , due to the Lipschitz continuous assumption on @math . Note that in our main template , @math , which is not Lipschitz continuous. | {
"cite_N": [
"@cite_15"
],
"mid": [
"2055400003"
],
"abstract": [
"In this paper we first study a smooth optimization approach for solving a class of nonsmooth strictly concave maximization problems whose objective functions admit smooth convex minimization reformulations. In particular, we apply Nesterov's smooth optimization technique [Y. E. Nesterov, Dokl. Akad. Nauk SSSR, 269 (1983), pp. 543-547; Y. E. Nesterov, Math. Programming, 103 (2005), pp. 127-152] to their dual counterparts that are smooth convex problems. It is shown that the resulting approach has @math iteration complexity for finding an @math -optimal solution to both primal and dual problems. We then discuss the application of this approach to sparse covariance selection that is approximately solved as an @math -norm penalized maximum likelihood estimation problem, and also propose a variant of this approach which has substantially outperformed the latter one in our computational experiments. We finally compare the performance of these approaches with other first-order methods, namely, Nesterov's @math smooth approximation scheme and block-coordinate descent method studied in [A. d'Aspremont, O. Banerjee, and L. El Ghaoui, SIAM J. Matrix Anal. Appl., 30 (2008), pp. 56-66; J. Friedman, T. Hastie, and R. Tibshirani, Biostatistics, 9 (2008), pp. 432-441] for sparse covariance selection on a set of randomly generated instances. It shows that our smooth optimization approach substantially outperforms the first method above, and moreover, its variant substantially outperforms both methods above."
]
} |
1902.00297 | 2914814133 | Higher inductive-inductive types (HIITs) generalize inductive types of dependent type theories in two ways. On the one hand they allow the simultaneous definition of multiple sorts that can be indexed over each other. On the other hand they support equality constructors, thus generalizing higher inductive types of homotopy type theory. Examples that make use of both features are the Cauchy real numbers and the well-typed syntax of type theory where conversion rules are given as equality constructors. In this paper we propose a general definition of HIITs using a small type theory, named the theory of signatures. A context in this theory encodes a HIIT by listing the constructors. We also compute notions of induction and recursion for HIITs, by using variants of syntactic logical relation translations. Building full categorical semantics and constructing initial algebras is left for future work. The theory of HIIT signatures was formalised in Agda together with the syntactic translations. We also provide a Haskell implementation, which takes signatures as input and outputs translation results as valid Agda code. | The article of @cite_12 gives specification and semantics of QIITs in a set-truncated setting. Signatures are given as lists of functors which can be interpreted as complete categories of algebras, and completeness is used to talk about notions of induction and recursion. However, no strict positivity restriction is given, nor a construction of initial algebras. | {
"cite_N": [
"@cite_12"
],
"mid": [
"2900009437"
],
"abstract": [
"Quotient inductive-inductive types (QIITs) generalise inductive types in two ways: a QIIT can have more than one sort and the later sorts can be indexed over the previous ones. In addition, equality constructors are also allowed. We work in a setting with uniqueness of identity proofs, hence we use the term QIIT instead of higher inductive-inductive type. An example of a QIIT is the well-typed (intrinsic) syntax of type theory quotiented by conversion. In this paper first we specify finitary QIITs using a domain-specific type theory which we call the theory of signatures. The syntax of the theory of signatures is given by a QIIT as well. Then, using this syntax we show that all specified QIITs exist and they have a dependent elimination principle. We also show that algebras of a signature form a category with families (CwF) and use the internal language of this CwF to show that dependent elimination is equivalent to initiality."
]
} |
1902.00297 | 2914814133 | Higher inductive-inductive types (HIITs) generalize inductive types of dependent type theories in two ways. On the one hand they allow the simultaneous definition of multiple sorts that can be indexed over each other. On the other hand they support equality constructors, thus generalizing higher inductive types of homotopy type theory. Examples that make use of both features are the Cauchy real numbers and the well-typed syntax of type theory where conversion rules are given as equality constructors. In this paper we propose a general definition of HIITs using a small type theory, named the theory of signatures. A context in this theory encodes a HIIT by listing the constructors. We also compute notions of induction and recursion for HIITs, by using variants of syntactic logical relation translations. Building full categorical semantics and constructing initial algebras is left for future work. The theory of HIIT signatures was formalised in Agda together with the syntactic translations. We also provide a Haskell implementation, which takes signatures as input and outputs translation results as valid Agda code. | Closely related to the current work is the paper by the current authors and Altenkirch @cite_1 , which also concerns QIITs. There, signatures for QIITs are essentially a restriction of the signatures given here, but in contrast to the current work, the restricted quotient setting enables building initial algebras and detailed categorical semantics. | {
"cite_N": [
"@cite_1"
],
"mid": [
"2900009437"
],
"abstract": [
"Quotient inductive-inductive types (QIITs) generalise inductive types in two ways: a QIIT can have more than one sort and the later sorts can be indexed over the previous ones. In addition, equality constructors are also allowed. We work in a setting with uniqueness of identity proofs, hence we use the term QIIT instead of higher inductive-inductive type. An example of a QIIT is the well-typed (intrinsic) syntax of type theory quotiented by conversion. In this paper first we specify finitary QIITs using a domain-specific type theory which we call the theory of signatures. The syntax of the theory of signatures is given by a QIIT as well. Then, using this syntax we show that all specified QIITs exist and they have a dependent elimination principle. We also show that algebras of a signature form a category with families (CwF) and use the internal language of this CwF to show that dependent elimination is equivalent to initiality."
]
} |
1902.00297 | 2914814133 | Higher inductive-inductive types (HIITs) generalize inductive types of dependent type theories in two ways. On the one hand they allow the simultaneous definition of multiple sorts that can be indexed over each other. On the other hand they support equality constructors, thus generalizing higher inductive types of homotopy type theory. Examples that make use of both features are the Cauchy real numbers and the well-typed syntax of type theory where conversion rules are given as equality constructors. In this paper we propose a general definition of HIITs using a small type theory, named the theory of signatures. A context in this theory encodes a HIIT by listing the constructors. We also compute notions of induction and recursion for HIITs, by using variants of syntactic logical relation translations. Building full categorical semantics and constructing initial algebras is left for future work. The theory of HIIT signatures was formalised in Agda together with the syntactic translations. We also provide a Haskell implementation, which takes signatures as input and outputs translation results as valid Agda code. | The logical predicate syntactic translation was introduced by @cite_13 . The idea that a context can be seen as a signatures and the logical predicate translation can be used to derive the types of induction motives and methods was described in [Section 5.3] ttintt . Logical relations are used to derive the computation rules in [Section 4.3] kaposi-phd , but only for closed QIITs. Syntactic translations in the context of the calculus of inductive constructions are discussed in @cite_16 . Logical relations and parametricity can also be used to justify the existence of inductive types in a type theory with an impredicative universe @cite_38 . | {
"cite_N": [
"@cite_38",
"@cite_16",
"@cite_13"
],
"mid": [
"1847820984",
"2565502105",
"1497884533"
],
"abstract": [
"The paper describes an extension of well-founded semantics for logic programs with two types of negation. In this extension information about preferences between rules can be expressed in the logical language and derived dynamically. This is achieved by using a reserved predicate symbol and a naming technique. Conflicts among rules are resolved whenever possible on the basis of derived preference information. The well-founded conclusions of prioritized logic programs can be computed in polynomial time. A legal reasoning example illustrates the usefulness of the approach.",
"A family of syntactic models for the calculus of construction with universes (CCω) is described, all of them preserving conversion of the calculus definitionally, and thus giving rise directly to a program transformation of CCω into itself. Those models are based on the remark that negative type constructors (e.g. dependent product, coinductive types or universes) are underspecified in type theory-which leaves some freedom on extra intensional specifications. The model construction can be seen as a compilation phase from a complex type theory into a simpler type theory. Such models can be used to derive (the negative part of) independence results with respect to CCω, such as functional extensionality, propositional extensionality, univalence or the fact that bisimulation on a coinductive type may not coincide with equality. They can also be used to add new principles to the theory, which we illustrate by defining a version of CCω with ad-hoc polymorphism that shows in particular that parametricity is not an implicit requirement of type theory. The correctness of some of the models program transformations have been checked in the Coq proof assistant and have been instrumented as a Coq plugin.",
"Techniques such as verification condition generation, predicate abstraction, and expressive type systems reduce software verification to proving formulas in expressive logics. Programs and their specifications often make use of data structures such as sets, multisets, algebraic data types, or graphs. Consequently, formulas generated from verification also involve such data structures. To automate the proofs of such formulas we propose a logic (a “calculus”) of such data structures. We build the calculus by starting from decidable logics of individual data structures, and connecting them through functions and sets, in ways that go beyond the frameworks such as Nelson-Oppen. The result are new decidable logics that can simultaneously specify properties of different kinds of data structures and overcome the limitations of the individual logics. Several of our decidable logics include abstraction functions that map a data structure into its more abstract view (a tree into a multiset, a multiset into a set), into a numerical quantity (the size or the height), or into the truth value of a candidate data structure invariant (sortedness, or the heap property). For algebraic data types, we identify an asymptotic many-to-one condition on the abstraction function that guarantees the existence of a decision procedure. In addition to the combination based on abstraction functions, we can combine multiple data structure theories if they all reduce to the same data structure logic. As an instance of this approach, we describe a decidable logic whose formulas are propositional combinations of formulas in: weak monadic second-order logic of two successors, two-variable logic with counting, multiset algebra with Presburger arithmetic, the Bernays-Schonfinkel-Ramsey class of first-order logic, and the logic of algebraic data types with the set content function. The subformulas in this combination can share common variables that refer to sets of objects along with the common set algebra operations. Such sound and complete combination is possible because the relations on sets definable in the component logics are all expressible in Boolean Algebra with Presburger Arithmetic. Presburger arithmetic and its new extensions play an important role in our decidability results. In several cases, when we combine logics that belong to NP, we can prove the satisfiability for the combined logic is still in NP."
]
} |
1902.00113 | 2913620370 | Domain generalization (DG) is the challenging and topical problem of learning models that generalize to novel testing domains with different statistics than a set of known training domains. The simple approach of aggregating data from all source domains and training a single deep neural network end-to-end on all the data provides a surprisingly strong baseline that surpasses many prior published methods. In this paper we build on this strong baseline by designing an episodic training procedure that trains a single deep network in a way that exposes it to the domain shift that characterises a novel domain at runtime. Specifically, we decompose a deep network into feature extractor and classifier components, and then train each component by simulating it interacting with a partner who is badly tuned for the current domain. This makes both components more robust, ultimately leading to our networks producing state-of-the-art performance on three DG benchmarks. Furthermore, we consider the pervasive workflow of using an ImageNet trained CNN as a fixed feature extractor for downstream recognition tasks. Using the Visual Decathlon benchmark, we demonstrate that our episodic-DG training improves the performance of such a general purpose feature extractor by explicitly training a feature for robustness to novel problems. This shows that DG training can benefit standard practice in computer vision. | Multi-Domain Learning (MDL) MDL aims to learn several domains simultaneously using a single model @cite_7 @cite_32 @cite_27 @cite_30 . Depending on the problem, how much data is available per domain, and how similar the domains are, multi-domain learning can improve @cite_30 -- or sometimes worsen @cite_7 @cite_32 @cite_27 -- performance compared to a single model per domain. MDL is related to DG because the typical setting for DG is to assume a similar setup in that multiple source domains are provided. But that now the goal is to learn how to extract a domain-agnostic or domain-robust model from all those source domains. The most rigorous benchmark for MDL is the Visual Decathlon (VD) @cite_32 . We repurpose this benchmark for DG by training a CNN on a subset of the VD domains, and then evaluating its performance as a feature extractor on an unseen disjoint subset of them. We are the first to demonstrate DG at this scale, and in the heterogeneous label setting required for VD. | {
"cite_N": [
"@cite_30",
"@cite_27",
"@cite_32",
"@cite_7"
],
"mid": [
"2785994509",
"2786081970",
"2557626841",
"2897482938"
],
"abstract": [
"Many text classification tasks are known to be highly domain-dependent. Unfortunately, the availability of training data can vary drastically across domains. Worse still, for some domains there may not be any annotated data at all. In this work, we propose a multinomial adversarial network (MAN) to tackle the text classification problem in this real-world multidomain setting (MDTC). We provide theoretical justifications for the MAN framework, proving that different instances of MANs are essentially minimizers of various f-divergence metrics (Ali and Silvey, 1966) among multiple probability distributions. MANs are thus a theoretically sound generalization of traditional adversarial networks that discriminate over two distributions. More specifically, for the MDTC task, MAN learns features that are invariant across multiple domains by resorting to its ability to reduce the divergence among the feature distributions of each domain. We present experimental results showing that MANs significantly outperform the prior art on the MDTC task. We also show that MANs achieve state-of-the-art performance for domains with no labeled data.",
"While domain adaptation has been actively researched in recent years, most theoretical results and algorithms focus on the single-source-single-target adaptation setting. Naive application of such algorithms on multiple source domain adaptation problem may lead to suboptimal solutions. We propose a new generalization bound for domain adaptation when there are multiple source domains with labeled instances and one target domain with unlabeled instances. Compared with existing bounds, the new bound does not require expert knowledge about the target distribution, nor the optimal combination rule for multisource domains. Interestingly, our theory also leads to an efficient learning strategy using adversarial neural networks: we show how to interpret it as learning feature representations that are invariant to the multiple domain shifts while still being discriminative for the learning task. To this end, we propose two models, both of which we call multisource domain adversarial networks (MDANs): the first model optimizes directly our bound, while the second model is a smoothed approximation of the first one, leading to a more data-efficient and task-adaptive model. The optimization tasks of both models are minimax saddle point problems that can be optimized by adversarial training. To demonstrate the effectiveness of MDANs, we conduct extensive experiments showing superior adaptation performance on three real-world datasets: sentiment analysis, digit classification, and vehicle counting.",
"In this paper, we propose an approach to the domain adaptation, dubbed Second-or Higher-order Transfer of Knowledge (So-HoT), based on the mixture of alignments of second-or higher-order scatter statistics between the source and target domains. The human ability to learn from few labeled samples is a recurring motivation in the literature for domain adaptation. Towards this end, we investigate the supervised target scenario for which few labeled target training samples per category exist. Specifically, we utilize two CNN streams: the source and target networks fused at the classifier level. Features from the fully connected layers fc7 of each network are used to compute second-or even higher-order scatter tensors, one per network stream per class. As the source and target distributions are somewhat different despite being related, we align the scatters of the two network streams of the same class (within-class scatters) to a desired degree with our bespoke loss while maintaining good separation of the between-class scatters. We train the entire network in end-to-end fashion. We provide evaluations on the standard Office benchmark (visual domains) and RGB-D combined with Caltech256 (depth-to-rgb transfer). We attain state-of-the-art results.",
"Domain adaptation aims to train a model on labeled data from a source domain while minimizing test error on a target domain. Most of existing domain adaptation methods only focus on reducing domain shift of single-modal data. In this paper, we consider a new problem of multimodal domain adaptation and propose a unified framework to solve it. The proposed multimodal domain adaptation neural networks(MDANN) consist of three important modules. (1) A covariant multimodal attention is designed to learn a common feature representation for multiple modalities. (2) A fusion module adaptively fuses attended features of different modalities. (3) Hybrid domain constraints are proposed to comprehensively learn domain-invariant features by constraining single modal features, fused features, and attention scores. Through jointly attending and fusing under an adversarial objective, the most discriminative and domain-adaptive parts of the features are adaptively fused together. Extensive experimental results on two real-world cross-domain applications (emotion recognition and cross-media retrieval) demonstrate the effectiveness of the proposed method."
]
} |
1902.00113 | 2913620370 | Domain generalization (DG) is the challenging and topical problem of learning models that generalize to novel testing domains with different statistics than a set of known training domains. The simple approach of aggregating data from all source domains and training a single deep neural network end-to-end on all the data provides a surprisingly strong baseline that surpasses many prior published methods. In this paper we build on this strong baseline by designing an episodic training procedure that trains a single deep network in a way that exposes it to the domain shift that characterises a novel domain at runtime. Specifically, we decompose a deep network into feature extractor and classifier components, and then train each component by simulating it interacting with a partner who is badly tuned for the current domain. This makes both components more robust, ultimately leading to our networks producing state-of-the-art performance on three DG benchmarks. Furthermore, we consider the pervasive workflow of using an ImageNet trained CNN as a fixed feature extractor for downstream recognition tasks. Using the Visual Decathlon benchmark, we demonstrate that our episodic-DG training improves the performance of such a general purpose feature extractor by explicitly training a feature for robustness to novel problems. This shows that DG training can benefit standard practice in computer vision. | Neural Network Meta-Learning Learning-to-learn and meta-learning methods have resurged recently, in particular in few-shot recognition @cite_16 @cite_39 @cite_8 , and learning-to-optimize @cite_34 tasks. Despite signifiant other differences in motivation and methodological formalisations, a common feature of these methods is episodic training strategy. In the case of few-shot learning, the intuition is that while lot of source tasks and data may be available, these should be used for training in a way so as to closely simulate the testing conditions. Therefore at each learning iteration, a random subset of source tasks and instances are sampled to generate a training episode defined by a random few-shot learning task of similar data volume and cardinality as the model is expected to be tested on at runtime. Thus the model eventually sees' all the training data in aggregate, but in any given iteration, it is evaluated in a condition similar to a real testing' condition. In this paper we aim to develop an episodic training strategy to improve domain-robustness, rather than learning-to-learn. While the high-level idea of an episodic strategy is the same, the DG problem and associated episode construction details are completely different. | {
"cite_N": [
"@cite_8",
"@cite_16",
"@cite_34",
"@cite_39"
],
"mid": [
"2903334135",
"2807352135",
"2964078140",
"2786928087"
],
"abstract": [
"Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. The key idea is to leverage a large number of similar few-shot tasks in order to learn how to adapt a base-learner to a new task for which only a few labeled samples are available. As deep neural networks (DNNs) tend to overfit using a few samples only, meta-learning typically uses shallow neural networks (SNNs), thus limiting its effectiveness. In this paper we propose a novel few-shot learning method called meta-transfer learning (MTL) which learns to adapt a deep NN for few shot learning tasks. Specifically, \"meta\" refers to training multiple tasks, and \"transfer\" is achieved by learning scaling and shifting functions of DNN weights for each task. In addition, we introduce the hard task (HT) meta-batch scheme as an effective learning curriculum for MTL. We conduct experiments using (5-class, 1-shot) and (5-class, 5-shot) recognition tasks on two challenging few-shot learning benchmarks: miniImageNet and Fewshot-CIFAR100. Extensive comparisons to related works validate that our meta-transfer learning approach trained with the proposed HT meta-batch scheme achieves top performance. An ablation study also shows that both components contribute to fast convergence and high accuracy.",
"Meta-learning for few-shot learning entails acquiring a prior over previous tasks and experiences, such that new tasks be learned from small amounts of data. However, a critical challenge in few-shot learning is task ambiguity: even when a powerful prior can be meta-learned from a large number of prior tasks, a small dataset for a new task can simply be too ambiguous to acquire a single model (e.g., a classifier) for that task that is accurate. In this paper, we propose a probabilistic meta-learning algorithm that can sample models for a new task from a model distribution. Our approach extends model-agnostic meta-learning, which adapts to new tasks via gradient descent, to incorporate a parameter distribution that is trained via a variational lower bound. At meta-test time, our algorithm adapts via a simple procedure that injects noise into gradient descent, and at meta-training time, the model is trained such that this stochastic adaptation procedure produces samples from the approximate model posterior. Our experimental results show that our method can sample plausible classifiers and regressors in ambiguous few-shot learning problems.",
"Meta-learning for few-shot learning entails acquiring a prior over previous tasks and experiences, such that new tasks be learned from small amounts of data. However, a critical challenge in few-shot learning is task ambiguity: even when a powerful prior can be meta-learned from a large number of prior tasks, a small dataset for a new task can simply be too ambiguous to acquire a single model (e.g., a classifier) for that task that is accurate. In this paper, we propose a probabilistic meta-learning algorithm that can sample models for a new task from a model distribution. Our approach extends model-agnostic meta-learning, which adapts to new tasks via gradient descent, to incorporate a parameter distribution that is trained via a variational lower bound. At meta-test time, our algorithm adapts via a simple procedure that injects noise into gradient descent, and at meta-training time, the model is trained such that this stochastic adaptation procedure produces samples from the approximate model posterior. Our experimental results show that our method can sample plausible classifiers and regressors in ambiguous few-shot learning problems.",
"Few-shot learning remains challenging for meta-learning that learns a learning algorithm (meta-learner) from many related tasks. In this work, we argue that this is due to the lack of a good representation for meta-learning, and propose deep meta-learning to integrate the representation power of deep learning into meta-learning. The framework is composed of three modules, a concept generator, a meta-learner, and a concept discriminator, which are learned jointly. The concept generator, e.g. a deep residual net, extracts a representation for each instance that captures its high-level concept, on which the meta-learner performs few-shot learning, and the concept discriminator recognizes the concepts. By learning to learn in the concept space rather than in the complicated instance space, deep meta-learning can substantially improve vanilla meta-learning, which is demonstrated on various few-shot image recognition problems. For example, on 5-way-1-shot image recognition on CIFAR-100 and CUB-200, it improves Matching Nets from 50.53 and 56.53 to 58.18 and 63.47 , improves MAML from 49.28 and 50.45 to 56.65 and 64.63 , and improves Meta-SGD from 53.83 and 53.34 to 61.62 and 66.95 , respectively."
]
} |
1902.00197 | 2913476238 | Monte Carlo (MC) permutation test is considered the gold standard for statistical hypothesis testing, especially when standard parametric assumptions are not clear or likely to fail. However, in modern data science settings where a large number of hypothesis tests need to be performed simultaneously, it is rarely used due to its prohibitive computational cost. In genome-wide association studies, for example, the number of hypothesis tests @math is around @math while the number of MC samples @math for each test could be greater than @math , totaling more than @math = @math samples. In this paper, we propose Adaptive MC multiple Testing (AMT) to estimate MC p-values and control false discovery rate in multiple testing. The algorithm outputs the same result as the standard full MC approach with high probability while requiring only @math samples. This sample complexity is shown to be optimal. On a Parkinson GWAS dataset, the algorithm reduces the running time from 2 months for full MC to an hour. The AMT algorithm is derived based on the theory of multi-armed bandits. | The problem of multiple testing with MC p-values has been studied in the broader statistical literature. Interesting heuristic adaptive algorithms were proposed without formal FDR guarantee @cite_29 @cite_41 ; the latter was developed via modifying Thompson sampling, another MAB algorithm. Asymptotic results were provided that the output of the adaptive algorithms will converge to the desired set of discoveries @cite_35 @cite_46 @cite_4 . Specifically, the most recent work @cite_4 provided a general result that incorporates virtually all popular multiple testing procedures. However, none of the above works provide a standard FDR control guarantee (e.g., @math ) nor an analysis of the MC sample complexity; the MC sample complexity was analyzed in another work only for the case of using Bonferroni procedure @cite_32 . In the present work, standard FDR control guarantee is provided, as well as both upper and lower bounds on the MC sample complexity, establishing the optimality of AMT . | {
"cite_N": [
"@cite_35",
"@cite_4",
"@cite_41",
"@cite_29",
"@cite_32",
"@cite_46"
],
"mid": [
"2099082142",
"1859329799",
"2002917158",
"1933042805",
"2406252837",
"1556276033"
],
"abstract": [
"Motivation: In molecular biology, as in many other scientific fields, the scale of analyses is ever increasing. Often, complex Monte Carlo simulation is required, sometimes within a large-scale multiple testing setting. The resulting computational costs may be prohibitively high. Results: We here present MCFDR, a simple, novel algorithm for false discovery rate (FDR) modulated sequential Monte Carlo (MC) multiple hypothesis testing. The algorithm iterates between adding MC samples across tests and calculating intermediate FDR values for the collection of tests. MC sampling is stopped either by sequential MC or based on a threshold on FDR. An essential property of the algorithm is that it limits the total number of MC samples whatever the number of true null hypotheses. We show on both real and simulated data that the proposed algorithm provides large gains in computational efficiency. Availability: MCFDR is implemented in the Genomic HyperBrowser ( http: hyperbrowser.uio.no mcfdr), a web-based system for genome analysis. All input data and results are available and can be reproduced through a Galaxy Pages document at: http: hyperbrowser.uio.no mcfdr u sandve p mcfdr. Contact: geirksa@ifi.uio.no",
"Multiple testing is often carried out in practice using approximated p-values obtained, for instance, via bootstrap or permutation tests. We are interested in allocating a pre-specified total number of samples (that is draws from a bootstrap distribution or permutations) to all hypotheses in order to approximate their p-values in an optimal way, in the sense that the allocation minimizes the total expected number of misclassified hypotheses. By a misclassified hypothesis we refer to a decision on single hypotheses which differs from the one obtained if all p-values were known analytically. Neither using a constant number of samples per p-value estimate nor more sophisticated approaches available in the literature guarantee the computation of an optimal allocation in the above sense. This article derives the optimal allocation of a finite total number of samples to a finite number of hypotheses tested using the Bonferroni correction. Simulation studies show that a simple sampling algorithm based on Thompson Sampling asympotically mimics this optimal allocation.",
"It is a common practice to use resampling methods such as the bootstrap for calculating the p-value for each test when performing large scale multiple testing. The precision of the bootstrap p-values and that of the false discovery rate (FDR) relies on the number of bootstraps used for testing each hypothesis. Clearly, the larger the number of bootstraps the better the precision. However, the required number of bootstraps can be computationally burdensome, and it multiplies the number of tests to be performed. Further adding to the computational challenge is that in some applications the calculation of the test statistic itself may require considerable computation time. As technology improves one can expect the dimension of the problem to increase as well. For instance, during the early days of microarray technology, the number of probes on a cDNA chip was less than 10,000. Now the Affymetrix chips come with over 50,000 probes per chip. Motivated by this important need, we developed a simple adaptive bootstrap methodology for large scale multiple testing, which reduces the total number of bootstrap calculations while ensuring the control of the FDR. The proposed algorithm results in a substantial reduction in the number of bootstrap samples. Based on a simulation study we found that, relative to the number of bootstraps required for the Benjamini-Hochberg (BH) procedure, the standard FDR methodology which was the proposed methodology achieved a very substantial reduction in the number of bootstraps. In some cases the new algorithm required as little as 1 6th the number of bootstraps as the conventional BH procedure. Thus, if the conventional BH procedure used 1,000 bootstraps, then the proposed method required only 160 bootstraps. This methodology has been implemented for time-course dose-response data in our software, ORIOGEN, which is available from the authors upon request.",
"We are concerned with a situation in which we would like to test multiple hypotheses with tests whose p-values cannot be computed explicitly but can be approximated using Monte Carlo simulation. This scenario occurs widely in practice. We are interested in obtaining the same rejections and non-rejections as the ones obtained if the p-values for all hypotheses had been available. The present article introduces a framework for this scenario by providing a generic algorithm for a general multiple testing procedure. We establish conditions that guarantee that the rejections and non-rejections obtained through Monte Carlo simulations are identical to the ones obtained with the p-values. Our framework is applicable to a general class of step-up and step-down procedures, which includes many established multiple testing corrections such as the ones of Bonferroni, Holm, Sidak, Hochberg or Benjamini–Hochberg. Moreover, we show how to use our framework to improve algorithms available in the literature in such a way as to yield theoretical guarantees on their results. These modifications can easily be implemented in practice and lead to a particular way of reporting multiple testing results as three sets together with an error bound on their correctness, demonstrated exemplarily using a real biological dataset.",
"We present a new Markov Chain Monte Carlo (MCMC) sampling algorithm for probabilistic programs. Our approach and tool, called R2, has the unique feature of employing program analysis in order to improve the efficiency of MCMC sampling. Given an input program P, R2 propagates observations in P backwards to obtain a semantically equivalent program P′ in which every probabilistic assignment is immediately followed by an observe statement. Inference is performed by a suitably modified version of the Metropolis-Hastings algorithm that exploits the structure of the program P′. This has the overall effect of preventing rejections due to program executions that fail to satisfy observations in P. We formalize the semantics of probabilistic programs and rigorously prove the correctness of R2. We also empirically demonstrate the effectiveness of R2--in particular, we show that R2 is able to produce results of similar quality as the CHURCH and STAN probabilistic programming tools with much shorter execution time.",
"In the conclusion of his monumental paper on optimal inapproximability results, Hastad [13] suggested that Fourier analysis of Dictator (Long Code) Tests may not be universally applicable in the study of CSPs. His main open question was to determine if the technique could resolve the approximability of satisfiable 3-bit constraint satisfaction problems. In particular, he asked if the \"Not Two\" (NTW) predicate is non-approximable beyond the random assignment threshold of 5 8 on satisfiable instances. Around the same time, Zwick [30] showed that all satisfiable 3-CSPs are 5 8-approximable and conjectured that the 5 8 is optimal. In this work we show that Fourier analysis techniques can produce a Dictator Test based on NTW with completeness 1 and soundness 5 8. Our test's analysis uses the Bonami-Gross-Beckner hypercontractive inequality. We also show a soundness lower bound of 5 8 for all 3-query Dictator Tests with perfect completeness. This lower bound for Property Testing is proved in part via a semidefinite programming algorithm of Zwick [30]. Our work precisely determines the 3-query \"Dictatorship Testing gap\". Although this represents progress on Zwick's conjecture, current PCP \"outer verifier\" technology is insufficient to convert our Dictator Test into an NP-hardness-of-approximation result."
]
} |
1902.00197 | 2913476238 | Monte Carlo (MC) permutation test is considered the gold standard for statistical hypothesis testing, especially when standard parametric assumptions are not clear or likely to fail. However, in modern data science settings where a large number of hypothesis tests need to be performed simultaneously, it is rarely used due to its prohibitive computational cost. In genome-wide association studies, for example, the number of hypothesis tests @math is around @math while the number of MC samples @math for each test could be greater than @math , totaling more than @math = @math samples. In this paper, we propose Adaptive MC multiple Testing (AMT) to estimate MC p-values and control false discovery rate in multiple testing. The algorithm outputs the same result as the standard full MC approach with high probability while requiring only @math samples. This sample complexity is shown to be optimal. On a Parkinson GWAS dataset, the algorithm reduces the running time from 2 months for full MC to an hour. The AMT algorithm is derived based on the theory of multi-armed bandits. | There have been works on fast permutation test for GWAS @cite_8 @cite_11 @cite_10 @cite_17 ; they consider a different goal which is to accelerate the process of separately computing each MC p-value. In contrast, AMT accelerates the entire workflow of both computing MC p-values and applying BH on them, where the decision for each hypothesis also depends globally on others. The state-of-art method is the sequential Monte Carlo procedure (sMC) that is implemented in the popular GWAS package PLINK @cite_49 @cite_44 @cite_39 . For each hypothesis, it keeps MC sampling until having observed @math extreme events or hit the sampling cap @math . Then BH is applied on the set of sMC p-values. Here we note that the sMC p-values are conservative so this procedure controls FDR. sMC is discussed and thoroughly compared against in the rest of the paper. | {
"cite_N": [
"@cite_8",
"@cite_17",
"@cite_39",
"@cite_44",
"@cite_49",
"@cite_10",
"@cite_11"
],
"mid": [
"2099082142",
"2093211258",
"1797876566",
"1859329799",
"2394967154",
"2124333705",
"2774890264"
],
"abstract": [
"Motivation: In molecular biology, as in many other scientific fields, the scale of analyses is ever increasing. Often, complex Monte Carlo simulation is required, sometimes within a large-scale multiple testing setting. The resulting computational costs may be prohibitively high. Results: We here present MCFDR, a simple, novel algorithm for false discovery rate (FDR) modulated sequential Monte Carlo (MC) multiple hypothesis testing. The algorithm iterates between adding MC samples across tests and calculating intermediate FDR values for the collection of tests. MC sampling is stopped either by sequential MC or based on a threshold on FDR. An essential property of the algorithm is that it limits the total number of MC samples whatever the number of true null hypotheses. We show on both real and simulated data that the proposed algorithm provides large gains in computational efficiency. Availability: MCFDR is implemented in the Genomic HyperBrowser ( http: hyperbrowser.uio.no mcfdr), a web-based system for genome analysis. All input data and results are available and can be reproduced through a Galaxy Pages document at: http: hyperbrowser.uio.no mcfdr u sandve p mcfdr. Contact: geirksa@ifi.uio.no",
"We present a method for controlling the output of procedural modeling programs using Sequential Monte Carlo (SMC). Previous probabilistic methods for controlling procedural models use Markov Chain Monte Carlo (MCMC), which receives control feedback only for completely-generated models. In contrast, SMC receives feedback incrementally on incomplete models, allowing it to reallocate computational resources and converge quickly. To handle the many possible sequentializations of a structured, recursive procedural modeling program, we develop and prove the correctness of a new SMC variant, Stochastically-Ordered Sequential Monte Carlo (SOSMC). We implement SOSMC for general-purpose programs using a new programming primitive: the stochastic future. Finally, we show that SOSMC reliably generates high-quality outputs for a variety of programs and control scoring functions. For small computational budgets, SOSMC's outputs often score nearly twice as high as those of MCMC or normal SMC.",
"Tasks such as record linkage and multi-target tracking, which involve reconstructing the set of objects that underlie some observed data, are particularly challenging for probabilistic inference. Recent work has achieved efficient and accurate inference on such problems using Markov chain Monte Carlo (MCMC) techniques with customized proposal distributions. Currently, implementing such a system requires coding MCMC state representations and acceptance probability calculations that are specific to a particular application. An alternative approach, which we pursue in this paper, is to use a general-purpose probabilistic modeling language (such as BLOG) and a generic Metropolis-Hastings MCMC algorithm that supports user-supplied proposal distributions. Our algorithm gains flexibility by using MCMC states that are only partial descriptions of possible worlds; we provide conditions under which MCMC over partial worlds yields correct answers to queries. We also show how to use a context-specific Bayes net to identify the factors in the acceptance probability that need to be computed for a given proposed move. Experimental results on a citation matching task show that our general-purpose MCMC engine compares favorably with an application-specific system.",
"Multiple testing is often carried out in practice using approximated p-values obtained, for instance, via bootstrap or permutation tests. We are interested in allocating a pre-specified total number of samples (that is draws from a bootstrap distribution or permutations) to all hypotheses in order to approximate their p-values in an optimal way, in the sense that the allocation minimizes the total expected number of misclassified hypotheses. By a misclassified hypothesis we refer to a decision on single hypotheses which differs from the one obtained if all p-values were known analytically. Neither using a constant number of samples per p-value estimate nor more sophisticated approaches available in the literature guarantee the computation of an optimal allocation in the above sense. This article derives the optimal allocation of a finite total number of samples to a finite number of hypotheses tested using the Bonferroni correction. Simulation studies show that a simple sampling algorithm based on Thompson Sampling asympotically mimics this optimal allocation.",
"Recently, Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC) methods have been proposed for scaling up Monte Carlo computations to large data problems. Whilst these approaches have proven useful in many applications, vanilla SG-MCMC might suffer from poor mixing rates when random variables exhibit strong couplings under the target densities or big scale differences. In this study, we propose a novel SG-MCMC method that takes the local geometry into account by using ideas from Quasi-Newton optimization methods. These second order methods directly approximate the inverse Hessian by using a limited history of samples and their gradients. Our method uses dense approximations of the inverse Hessian while keeping the time and memory complexities linear with the dimension of the problem. We provide a formal theoretical analysis where we show that the proposed method is asymptotically unbiased and consistent with the posterior expectations. We illustrate the effectiveness of the approach on both synthetic and real datasets. Our experiments on two challenging applications show that our method achieves fast convergence rates similar to Riemannian approaches while at the same time having low computational requirements similar to diagonal preconditioning approaches.",
"We propose sequential Monte Carlo (SMC) methods for sampling the posterior distribution of state-space models under highly informative observation regimes, a situation in which standard SMC methods can perform poorly. A special case is simulating bridges between given initial and final values. The basic idea is to introduce a schedule of intermediate weighting and resampling times between observation times, which guide particles towards the final state. This can always be done for continuous-time models, and may be done for discrete-time models under sparse observation regimes; our main focus is on continuous-time diffusion processes. The methods are broadly applicable in that they support multivariate models with partial observation, do not require simulation of the backward transition (which is often unavailable), and, where possible, avoid pointwise evaluation of the forward transition. When simulating bridges, the last characteristic cannot be avoided entirely without concessions, and we suggest an $ ...",
"In human-robot collaborative tasks, incorporating qualitative information provided by humans can greatly enhance the robustness and efficacy of robot state estimation. We introduce an algorithmic framework to model qualitative information as quantitative constraints on and between states. Our approach, named Sequentially Constrained Hamiltonian Monte Carlo, integrates Hamiltonian dynamics into Sequentially Constrained Monte Carlo sampling. We are able to generate samples that satisfy arbitrarily complex, non-smooth and discontinuous constraints, which in turn allows us to support a wide range of qualitative information. We evaluate our approach for constrained sampling qualitatively and quantitatively with several classes of constraints. SCHMC significantly outperforms the Metropolis-Hastings algorithm (a standard Markov Chain Monte Carlo (MCMC) method) and the Hamiltonian Monte Carlo (HMC) method, in terms of both the accuracy of the sampling (for satisfying constraints) and the quality of approximation. Compared to Sequentially Constrained Monte Carlo (SCMC), which supports similar kinds of constraints, our SCHMC approach has faster convergence rates and lower parameter sensitivity."
]
} |
1902.00127 | 2912839841 | Mixed datasets consist of both numeric and categorical attributes. Various K-means-based clustering algorithms have been developed to cluster these datasets. Generally, these algorithms use random partition as a starting point, which tend to produce different clustering results in different runs. This inconsistency of clustering results may lead to unreliable inferences from the data. A few initialization algorithms have been developed to compute initial partition for mixed datasets; however, they are either computationally expensive or they do not produce consistent clustering results in different runs. In this paper, we propose, initKmix, a novel approach to find initial partition for K-means-based clustering algorithms for mixed datasets. The initKmix is based on the experimental observations that (i) some data points in a dataset remain in the same clusters created by k-means-based clustering algorithm irrespective of the choice of initial clusters, and (ii) individual attribute information can be used to create initial clusters. In initKmix method, a k-means-based clustering algorithm is run many times, in each run one of the attribute is used to produce initial partition. The clustering results of various runs are combined to produce initial partition. This initial partition is then be used as a seed to a k-means-based clustering algorithm to cluster mixed data. The initial partitions produced by initKmix are always fixed, do not change over different runs or by changing the order of the data objects. Experiments with various categorical and mixed datasets showed that initKmix produced accurate and consistent results, and outperformed random initialization and other state-of-the-art initialization methods. Experiments also showed that K-means-based clustering for mixed datasets with initKmix outperformed many state-of-the-art clustering algorithms. | K-means clustering algorithm is a popular clustering algorithm for datasets consisting of numeric attributes because of its low computational complexity @cite_10 . The complexity is linear with respect to the number of data points and scales well for large datasets. It minimizes the optimization function presented in Equation 1 iteratively, | {
"cite_N": [
"@cite_10"
],
"mid": [
"2103660993"
],
"abstract": [
"K-means, a simple and effective clustering algorithm, is one of the most widely used algorithms in computer vision community. Traditional k-means is an iterative algorithm — in each iteration new cluster centers are computed and each data point is re-assigned to its nearest center. The cluster re-assignment step becomes prohibitively expensive when the number of data points and cluster centers are large. In this paper, we propose a novel approximate k-means algorithm to greatly reduce the computational complexity in the assignment step. Our approach is motivated by the observation that most active points changing their cluster assignments at each iteration are located on or near cluster boundaries. The idea is to efficiently identify those active points by pre-assembling the data into groups of neighboring points using multiple random spatial partition trees, and to use the neighborhood information to construct a closure for each cluster, in such a way only a small number of cluster candidates need to be considered when assigning a data point to its nearest cluster. Using complexity analysis, real data clustering, and applications to image retrieval, we show that our approach out-performs state-of-the-art approximate k-means algorithms in terms of clustering quality and efficiency."
]
} |
1902.00127 | 2912839841 | Mixed datasets consist of both numeric and categorical attributes. Various K-means-based clustering algorithms have been developed to cluster these datasets. Generally, these algorithms use random partition as a starting point, which tend to produce different clustering results in different runs. This inconsistency of clustering results may lead to unreliable inferences from the data. A few initialization algorithms have been developed to compute initial partition for mixed datasets; however, they are either computationally expensive or they do not produce consistent clustering results in different runs. In this paper, we propose, initKmix, a novel approach to find initial partition for K-means-based clustering algorithms for mixed datasets. The initKmix is based on the experimental observations that (i) some data points in a dataset remain in the same clusters created by k-means-based clustering algorithm irrespective of the choice of initial clusters, and (ii) individual attribute information can be used to create initial clusters. In initKmix method, a k-means-based clustering algorithm is run many times, in each run one of the attribute is used to produce initial partition. The clustering results of various runs are combined to produce initial partition. This initial partition is then be used as a seed to a k-means-based clustering algorithm to cluster mixed data. The initial partitions produced by initKmix are always fixed, do not change over different runs or by changing the order of the data objects. Experiments with various categorical and mixed datasets showed that initKmix produced accurate and consistent results, and outperformed random initialization and other state-of-the-art initialization methods. Experiments also showed that K-means-based clustering for mixed datasets with initKmix outperformed many state-of-the-art clustering algorithms. | k-Harmonic means clustering algorithm addresses the random initial clusters problem by using a different cost function @cite_17 for numeric datasets. K-Harmonic means clustering algorithm clusters create more stable clusters as compared to K-means clustering algorithm with random initial clusters. Ahmad and Hashmi @cite_15 combine the distance measure and the definition of cluster centres for mixed datasets suggested by Ahmad and Dey @cite_28 with k-Harmonic clustering algorithm @cite_17 to develop the k-Harmonic clustering algorithm for mixed datasets. Their method is less sensitive to the choice of initial cluster centres. The standard deviation of clustering accuracy of this method is small as compared to the random initialization method. However, it does not give the same clustering results in different runs with different initial partitions. | {
"cite_N": [
"@cite_28",
"@cite_15",
"@cite_17"
],
"mid": [
"2470462496",
"150776092",
"1970820654"
],
"abstract": [
"Display Omitted A K-Harmonic clustering algorithm for mixed data has been presented to reduce random initialization problem for partitional algorithms.The proposed clustering algorithm uses a distance measure developed for mixed datasets.The experiment results suggest that clustering results are quite insensitive to random initialization.The proposed algorithm performed better than other clustering algorithms for various datasets. K-means type clustering algorithms for mixed data that consists of numeric and categorical attributes suffer from cluster center initialization problem. The final clustering results depend upon the initial cluster centers. Random cluster center initialization is a popular initialization technique. However, clustering results are not consistent with different cluster center initializations. K-Harmonic means clustering algorithm tries to overcome this problem for pure numeric data. In this paper, we extend the K-Harmonic means clustering algorithm for mixed datasets. We propose a definition for a cluster center and a distance measure. These cluster centers and the distance measure are used with the cost function of K-Harmonic means clustering algorithm in the proposed algorithm. Experiments were carried out with pure categorical datasets and mixed datasets. Results suggest that the proposed clustering algorithm is quite insensitive to the cluster center initialization problem. Comparative studies with other clustering algorithms show that the proposed algorithm produce better clustering results.",
"We propose a new class of center-based iterative clustering algorithms, K-Harmonic Means (KHMp), which is essentially insensitive to the initialization of the centers, demonstrated through many experiments. The insensitivity to initialization is attributed to a dynamic weighting function, which increases the importance of the data points that are far from any centers in the next iteration. The dependency of the K-Means’ and EM’s performance on the initialization of the centers has been a major problem. Many have tried to generate good initializations to solve the sensitivity problem. KHMp addresses the intrinsic problem by replacing the minimum distance from a data point to the centers, used in K-Means, by the Harmonic Averages of the distances from the data point to all centers. KHMp significantly improves the quality of clustering results comparing with both K-Means and EM. The KHMp algorithms have been implemented in both sequential and parallel languages and tested on hundreds of randomly generated datasets with different data distribution and clustering characteristics.",
"In this paper we present a new clustering method based on k-means that have avoided alternative randomness of initial center. This paper focused on K-means algorithm to the initial value of the dependence of k selected from the aspects of the algorithm is improved. First,the initial clustering number is. Second, through the application of the sub-merger strategy the categories were combined.The algorithm does not require the user is given in advance the number of cluster. Experiments on synthetic datasets are presented to have shown significant improvements in clustering accuracy in comparison with the random k-means."
]
} |
1902.00245 | 2914121279 | Typical recommender systems push K items at once in the result page in the form of a feed, in which the selection and the order of the items are important for user experience. In this paper, we formalize the K-item recommendation problem as taking an unordered set of candidate items as input, and exporting an ordered list of selected items as output. The goal is to maximize the overall utility, e.g. the click through rate, of the whole list. As one solution to the K-item recommendation problem under this proposition, we proposed a new ranking framework called the Evaluator-Generator framework. In this framework, the Evaluator is trained on user logs to precisely predict the expected feedback of each item by fully considering its intra-list correlations with other co-exposed items. On the other hand, the Generator will generate different sequences from which the Evaluator will choose one sequence as the final recommendation. In our experiments, both the offline analysis and the online test show the effectiveness of our proposed framework. Furthermore, we show that the offline behavior of the Evaluator is consistent with the realistic online environment. | The position bias is also an important and practical issue in RS. Click models are widely studied in Information Retrieval systems, such as the Cascade Click Model( @cite_29 ) and the Dynamic Bayesian Network( @cite_8 ). It is found that the position bias is not only related to the user's personal habit, but also related to the layout design etc( @cite_34 ). Thus click models often need to be considered case-by-case. Few of the previous works have studied the intra-list correlation and the position bias together. Set based recommendation algorithms including submodular ranking and DPP do not consider the position bias at all. In contrast, our framework has naturally addressed the position bias and intra-list correlations altogether. | {
"cite_N": [
"@cite_29",
"@cite_34",
"@cite_8"
],
"mid": [
"1992549066",
"2099213975",
"2171357718"
],
"abstract": [
"Search engine click logs provide an invaluable source of relevance information, but this information is biased. A key source of bias is presentation order: the probability of click is influenced by a document's position in the results page. This paper focuses on explaining that bias, modelling how probability of click depends on position. We propose four simple hypotheses about how position bias might arise. We carry out a large data-gathering effort, where we perturb the ranking of a major search engine, to see how clicks are affected. We then explore which of the four hypotheses best explains the real-world position effects, and compare these to a simple logistic regression model. The data are not well explained by simple position models, where some users click indiscriminately on rank 1 or there is a simple decay of attention over ranks. A 'cascade' model, where users view results from top to bottom and leave as soon as they see a worthwhile document, is our best explanation for position bias in early ranks",
"As with any application of machine learning, web search ranking requires labeled data. The labels usually come in the form of relevance assessments made by editors. Click logs can also provide an important source of implicit feedback and can be used as a cheap proxy for editorial labels. The main difficulty however comes from the so called position bias - urls appearing in lower positions are less likely to be clicked even if they are relevant. In this paper, we propose a Dynamic Bayesian Network which aims at providing us with unbiased estimation of the relevance from the click logs. Experiments show that the proposed click model outperforms other existing click models in predicting both click-through rate and relevance.",
"Recent works in Recommender Systems (RS) have investigated the relationships between the prediction accuracy, i.e. the ability of a RS to minimize a cost function (for instance the RMSE measure) in estimating users’ preferences, and the accuracy of the recommendation list provided to users. State-of-the-art recommendation algorithms, which focus on the minimization of RMSE, have shown to achieve weak results from the recommendation accuracy perspective, and vice versa. In this work we present a novel Bayesian probabilistic hierarchical approach for users’ preference data, which is designed to overcome the limitation of current methodologies and thus to meet both prediction and recommendation accuracy. According to the generative semantics of this technique, each user is modeled as a random mixture over latent factors, which identify users community interests. Each individual user community is then modeled as a mixture of topics, which capture the preferences of the members on a set of items. We provide two dierent formalization of the basic hierarchical model: BH-Forced focuses on rating prediction, while BH-Free models both the popularity of items and the distribution over item ratings. The combined modeling of item popularity and rating provides a powerful framework for the generation of highly accurate recommendations. An extensive evaluation over two popular benchmark datasets reveals the eectiveness and the quality of the proposed algorithms, showing that BH-Free realizes the most satisfactory compromise between prediction and recommendation accuracy with respect to several stateof-the-art competitors."
]
} |
1902.00245 | 2914121279 | Typical recommender systems push K items at once in the result page in the form of a feed, in which the selection and the order of the items are important for user experience. In this paper, we formalize the K-item recommendation problem as taking an unordered set of candidate items as input, and exporting an ordered list of selected items as output. The goal is to maximize the overall utility, e.g. the click through rate, of the whole list. As one solution to the K-item recommendation problem under this proposition, we proposed a new ranking framework called the Evaluator-Generator framework. In this framework, the Evaluator is trained on user logs to precisely predict the expected feedback of each item by fully considering its intra-list correlations with other co-exposed items. On the other hand, the Generator will generate different sequences from which the Evaluator will choose one sequence as the final recommendation. In our experiments, both the offline analysis and the online test show the effectiveness of our proposed framework. Furthermore, we show that the offline behavior of the Evaluator is consistent with the realistic online environment. | We also borrow ideas from applying Reinforcement Learning to Combinatorial Optimization(CO) problems. The pointer network @cite_31 is proposed as neural tool for universal CO. @cite_15 further extended pointer-network by using policy gradients for learning. @cite_35 applied Q-Learning to CO problems on graphs. The aforementioned works have been focused on universal CO problems including Travelling Salesman Problem, Minimum Vertex Cover etc. | {
"cite_N": [
"@cite_35",
"@cite_31",
"@cite_15"
],
"mid": [
"2790770538",
"2786995169",
"2146989110"
],
"abstract": [
"We propose a framework for solving combinatorial optimization problems of which the output can be represented as a sequence of input elements. As an alternative to the Pointer Network, we parameterize a policy by a model based entirely on (graph) attention layers, and train it efficiently using REINFORCE with a simple and robust baseline based on a deterministic (greedy) rollout of the best policy found during training. We significantly improve over state-of-the-art results for learning algorithms for the 2D Euclidean TSP, reducing the optimality gap for a single tour construction by more than 75 (to 0.33 ) and 50 (to 2.28 ) for instances with 20 and 50 nodes respectively.",
"Many recent state-of-the-art recommender systems such as D-ATT, TransNet and DeepCoNN exploit reviews for representation learning. This paper proposes a new neural architecture for recommendation with reviews. Our model operates on a multi-hierarchical paradigm and is based on the intuition that not all reviews are created equal, i.e., only a selected few are important. The importance, however, should be dynamically inferred depending on the current target. To this end, we propose a review-by-review pointer-based learning scheme that extracts important reviews from user and item reviews and subsequently matches them in a word-by-word fashion. This enables not only the most informative reviews to be utilized for prediction but also a deeper word-level interaction. Our pointer-based method operates with a gumbel-softmax based pointer mechanism that enables the incorporation of discrete vectors within differentiable neural architectures. Our pointer mechanism is co-attentive in nature, learning pointers which are co-dependent on user-item relationships. Finally, we propose a multi-pointer learning scheme that learns to combine multiple views of user-item interactions. We demonstrate the effectiveness of our proposed model via extensive experiments on 24 benchmark datasets from Amazon and Yelp. Empirical results show that our approach significantly outperforms existing state-of-the-art models, with up to 19 and 71 relative improvement when compared to TransNet and DeepCoNN respectively. We study the behavior of our multi-pointer learning mechanism, shedding light on 'evidence aggregation' patterns in review-based recommender systems.",
"A central challenge to many fields of science and engineering involves minimizing non-convex error functions over continuous, high dimensional spaces. Gradient descent or quasi-Newton methods are almost ubiquitously used to perform such minimizations, and it is often thought that a main source of difficulty for these local methods to find the global minimum is the proliferation of local minima with much higher error than the global minimum. Here we argue, based on results from statistical physics, random matrix theory, neural network theory, and empirical evidence, that a deeper and more profound difficulty originates from the proliferation of saddle points, not local minima, especially in high dimensional problems of practical interest. Such saddle points are surrounded by high error plateaus that can dramatically slow down learning, and give the illusory impression of the existence of a local minimum. Motivated by these arguments, we propose a new approach to second-order optimization, the saddle-free Newton method, that can rapidly escape high dimensional saddle points, unlike gradient descent and quasi-Newton methods. We apply this algorithm to deep or recurrent neural network training, and provide numerical evidence for its superior optimization performance."
]
} |
1902.00245 | 2914121279 | Typical recommender systems push K items at once in the result page in the form of a feed, in which the selection and the order of the items are important for user experience. In this paper, we formalize the K-item recommendation problem as taking an unordered set of candidate items as input, and exporting an ordered list of selected items as output. The goal is to maximize the overall utility, e.g. the click through rate, of the whole list. As one solution to the K-item recommendation problem under this proposition, we proposed a new ranking framework called the Evaluator-Generator framework. In this framework, the Evaluator is trained on user logs to precisely predict the expected feedback of each item by fully considering its intra-list correlations with other co-exposed items. On the other hand, the Generator will generate different sequences from which the Evaluator will choose one sequence as the final recommendation. In our experiments, both the offline analysis and the online test show the effectiveness of our proposed framework. Furthermore, we show that the offline behavior of the Evaluator is consistent with the realistic online environment. | Some recent works have been addresses long-term rewards in recommender systems. @cite_33 also work on intra-list correlations, which applies policy gradient and Monte Carlo Tree Search(MCTS) to optimize the @math -NDCG in diversified ranking for the global optimum. Other works pursues long-term rewards in inter-list recommendations ( @cite_10 , @cite_2 ). Though @cite_10 also proposed treating a page of recommendation list as a whole, the intra-list correlations are not well modelled and analyzed in their work. Also their work is not testified on realistic online RS. In this paper, we more thoroughly investigates the general form of the intra-list correlations. We formalize the utility of the whole list as our final target, but we have also used item-level feedbacks. This has not been sufficiently studied in the above works. | {
"cite_N": [
"@cite_10",
"@cite_33",
"@cite_2"
],
"mid": [
"1797876566",
"2155912844",
"2042281163"
],
"abstract": [
"Tasks such as record linkage and multi-target tracking, which involve reconstructing the set of objects that underlie some observed data, are particularly challenging for probabilistic inference. Recent work has achieved efficient and accurate inference on such problems using Markov chain Monte Carlo (MCMC) techniques with customized proposal distributions. Currently, implementing such a system requires coding MCMC state representations and acceptance probability calculations that are specific to a particular application. An alternative approach, which we pursue in this paper, is to use a general-purpose probabilistic modeling language (such as BLOG) and a generic Metropolis-Hastings MCMC algorithm that supports user-supplied proposal distributions. Our algorithm gains flexibility by using MCMC states that are only partial descriptions of possible worlds; we provide conditions under which MCMC over partial worlds yields correct answers to queries. We also show how to use a context-specific Bayes net to identify the factors in the acceptance probability that need to be computed for a given proposed move. Experimental results on a citation matching task show that our general-purpose MCMC engine compares favorably with an application-specific system.",
"In this work we present topic diversification, a novel method designed to balance and diversify personalized recommendation lists in order to reflect the user's complete spectrum of interests. Though being detrimental to average accuracy, we show that our method improves user satisfaction with recommendation lists, in particular for lists generated using the common item-based collaborative filtering algorithm.Our work builds upon prior research on recommender systems, looking at properties of recommendation lists as entities in their own right rather than specifically focusing on the accuracy of individual recommendations. We introduce the intra-list similarity metric to assess the topical diversity of recommendation lists and the topic diversification approach for decreasing the intra-list similarity. We evaluate our method using book recommendation data, including offline analysis on 361, !, 349 ratings and an online study involving more than 2, !, 100 subjects.",
"Recommender systems apply knowledge discovery techniques to the problem of making personalized recommendations for information, products or services during a live interaction. These systems, especially the k-nearest neighbor collaborative ltering based ones, are achieving widespread success on the Web. The tremendous growth in the amount of available information and the number of visitors to Web sites in recent years poses some key challenges for recommender systems. These are: producing high quality recommendations, performing many recommendations per second for millions of users and items and achieving high coverage in the face of data sparsity. In traditional collaborative ltering systems the amount of work increases with the number of participants in the system. New recommender system technologies are needed that can quickly produce high quality recommendations, even for very large-scale problems. To address these issues we have explored item-based collaborative ltering techniques. Item-based techniques rst analyze the user-item matrix to identify relationships between di erent items, and then use these relationships to indirectly compute recommendations for users. In this paper we analyze di erent item-based recommendation generation algorithms. We look into di erent techniques for computing item-item similarities (e.g., item-item correlation vs. cosine similarities between item vectors) and di erent techniques for obtaining recommendations from them (e.g., weighted sum vs. regression model). Finally, we experimentally evaluate our results and compare them to the basic k-nearest neighbor approach. Our experiments suggest that item-based algorithms provide dramatically better performance than user-based algorithms, while at the same time providing better quality than the best available userbased algorithms."
]
} |
1901.11260 | 2912134658 | Many systems have to be maintained while the underlying constraints, costs and or profits change over time. Although the state of a system may evolve during time, a non-negligible transition cost is incured for transitioning from one state to another. In order to model such situations, (ICALP 2014) and (ICALP 2014) introduced a multistage model where the input is a sequence of instances (one for each time step), and the goal is to find a sequence of solutions (one for each time step) that are both (i) near optimal for each time step and (ii) as stable as possible. We focus on the multistage version of the Knapsack problem where we are given a time horizon t=1,2,...,T, and a sequence of knapsack instances I_1,I_2,...,I_T, one for each time step, defined on a set of n objects. In every time step t we have to choose a feasible knapsack S_t of I_t, which gives a knapsack profit. To measure the stability similarity of two consecutive solutions S_t and S_ t+1 , we identify the objects for which the decision, to be picked or not, remains the same in S_t and S_ t+1 , giving a transition profit. We are asked to produce a sequence of solutions S_1,S_2,...,S_T so that the total knapsack profit plus the overall transition profit is maximized. We propose a PTAS for the Multistage Knapsack problem. Then, we prove that there is no FPTAS for the problem even in the case where T=2, unless P=NP. Furthermore, we give a pseudopolynomial time algorithm for the case where the number of steps is bounded by a fixed constant and we show that otherwise the problem remains NP-hard even in the case where all the weights, profits and capacities are 0 or 1. | It is well known that for the usual Knapsack problem, in the continuous relaxation (variables in @math ), at most one variable is fractional. @cite_11 showed that this can be generalized for @math . | {
"cite_N": [
"@cite_11"
],
"mid": [
"1989921461"
],
"abstract": [
"Given @math elements with non-negative integer weights @math and an integer capacity @math , we consider the counting version of the classic knapsack problem: find the number of distinct subsets whose weights add up to at most @math . We give the first deterministic, fully polynomial-time approximation scheme (FPTAS) for estimating the number of solutions to any knapsack constraint (our estimate has relative error @math ). Our algorithm is based on dynamic programming. Previously, randomized polynomial-time approximation schemes (FPRAS) were known first by Morris and Sinclair via Markov chain Monte Carlo techniques, and subsequently by Dyer via dynamic programming and rejection sampling. In addition, we present a new method for deterministic approximate counting using read-once branching programs. Our approach yields an FPTAS for several other counting problems, including counting solutions for the multidimensional knapsack problem with a constant number of constraints, the general integer knapsack problem, and the contingency tables problem with a constant number of rows."
]
} |
1901.11260 | 2912134658 | Many systems have to be maintained while the underlying constraints, costs and or profits change over time. Although the state of a system may evolve during time, a non-negligible transition cost is incured for transitioning from one state to another. In order to model such situations, (ICALP 2014) and (ICALP 2014) introduced a multistage model where the input is a sequence of instances (one for each time step), and the goal is to find a sequence of solutions (one for each time step) that are both (i) near optimal for each time step and (ii) as stable as possible. We focus on the multistage version of the Knapsack problem where we are given a time horizon t=1,2,...,T, and a sequence of knapsack instances I_1,I_2,...,I_T, one for each time step, defined on a set of n objects. In every time step t we have to choose a feasible knapsack S_t of I_t, which gives a knapsack profit. To measure the stability similarity of two consecutive solutions S_t and S_ t+1 , we identify the objects for which the decision, to be picked or not, remains the same in S_t and S_ t+1 , giving a transition profit. We are asked to produce a sequence of solutions S_1,S_2,...,S_T so that the total knapsack profit plus the overall transition profit is maximized. We propose a PTAS for the Multistage Knapsack problem. Then, we prove that there is no FPTAS for the problem even in the case where T=2, unless P=NP. Furthermore, we give a pseudopolynomial time algorithm for the case where the number of steps is bounded by a fixed constant and we show that otherwise the problem remains NP-hard even in the case where all the weights, profits and capacities are 0 or 1. | @cite_11 use the result of Theorem to show that for any fixed constant @math @math admits a polynomial time approximation scheme (PTAS). Other PTASes have been presented in @cite_18 @cite_15 . Korte and Schrader @cite_6 showed that there is no FPTAS for @math unless @math . | {
"cite_N": [
"@cite_18",
"@cite_15",
"@cite_6",
"@cite_11"
],
"mid": [
"2567804199",
"2963006872",
"1972853793",
"2091602684"
],
"abstract": [
"An important question in theoretical computer science is to determine the best possible running time for solving a problem at hand. For geometric optimization problems, we often understand their complexity on a rough scale, but not very well on a finer scale. One such example is the two-dimensional knapsack problem for squares. There is a polynomial time (1 + ϵ)-approximation algorithm for it (i.e., a PTAS) but the running time of this algorithm is triple exponential in 1 ϵ, i.e., Ω(n221 ϵ). A double or triple exponential dependence on 1 ϵ is inherent in how this and several other algorithms for other geometric problems work. In this paper, we present an EPTAS for knapsack for squares, i.e., a (1+ϵ)-approximation algorithm with a running time of Oϵ(1)·nO(1). In particular, the exponent of n in the running time does not depend on ϵ at all! Since there can be no FPTAS for the problem (unless P = NP) this is the best kind of approximation scheme we can hope for. To achieve this improvement, we introduce two new key ideas: We present a fast method to guess the Ω(221 ϵ) relatively large squares of a suitable near-optimal packing instead of using brute-force enumeration. Secondly, we introduce an indirect guessing framework to define sizes of cells for the remaining squares. In the previous PTAS each of these steps needs a running time of Ω(n221 ϵ) and we improve both to Oϵ(1) · nO(1). We complete our result by giving an algorithm for two-dimensional knapsack for rectangles under (1 + ϵ)-resource augmentation. In this setting, we also improve the best known running time of Ω(n1 ϵ1 ϵ) to Oϵ(1) · nO(1) and compute even a solution with optimal profit, in contrast to the best previously known polynomial time algorithm for this setting that computes only an approximation. We believe that our new techniques have the potential to be useful for other settings as well.",
"In the Closest String problem one is given a family S of equal-length strings over some fixed alphabet, and the task is to find a string y that minimizes the maximum Hamming distance between y and a string from S. While polynomial-time approximation schemes (PTASes) for this problem are known for a long time [; J. ACM'02], no efficient polynomial-time approximation scheme (EPTAS) has been proposed so far. In this paper, we prove that the existence of an EPTAS for Closest String is in fact unlikely, as it would imply that FPT=W[1], a highly unexpected collapse in the hierarchy of parameterized complexity classes. Our proof also shows that the existence of a PTAS for Closest String with running time f(eps) n^o(1 eps), for any computable function f, would contradict the Exponential Time Hypothesis.",
"A polynomial time approximation scheme (PTAS) for an optimization problem A is an algorithm that given in input an instance of A and e > 0 finds a (1 + e)-approximate solution in time that is polynomial for each fixed e. Typical running times are nO(1e) or 21eO(1) n. While algorithms of the former kind tend to be impractical, the latter ones are more interesting. In several cases, the development of algorithms of the second type required considerably new, and sometimes harder, techniques. For some interesting problems, only nO(1e) approximation schemes are known. Under likely assumptions, we prove that for some problems (including natural ones) there cannot be approximation schemes running in time f(1 ϵ) n p0(1), no matter how fast function f grows. Our result relies on a connection with Parameterized Complexity Theory, and we show that this connection is necessary.",
"Assuming that NP @math @math BPTIME( @math ), we show that graph min-bisection, dense @math -subgraph, and bipartite clique have no polynomial time approximation scheme (PTAS). We give a reduction from the minimum distance of code (MDC) problem. Starting with an instance of MDC, we build a quasi-random probabilistically checkable proof (PCP) that suffices to prove the desired inapproximability results. In a quasi-random PCP, the query pattern of the verifier looks random in a certain precise sense. Among the several new techniques we introduce, the most interesting one gives a way of certifying that a given polynomial belongs to a given linear subspace of polynomials. As is important for our purpose, the certificate itself happens to be another polynomial, and it can be checked probabilistically by reading a constant number of its values."
]
} |
1901.11383 | 2952973608 | One of the most common modes of representing engineering schematics are Piping and Instrumentation diagrams (P&IDs) that describe the layout of an engineering process flow along with the interconnected process equipment. Over the years, P&ID diagrams have been manually generated, scanned and stored as image files. These files need to be digitized for purposes of inventory management and updation, and easy reference to different components of the schematics. There are several challenging vision problems associated with digitizing real world P&ID diagrams. Real world P&IDs come in several different resolutions, and often contain noisy textual information. Extraction of instrumentation information from these diagrams involves accurate detection of symbols that frequently have minute visual differences between them. Identification of pipelines that may converge and diverge at different points in the image is a further cause for concern. Due to these reasons, to the best of our knowledge, no system has been proposed for end-to-end data extraction from P&ID diagrams. However, with the advent of deep learning and the spectacular successes it has achieved in vision, we hypothesized that it is now possible to re-examine this problem armed with the latest deep learning models. To that end, we present a novel pipeline for information extraction from P&ID sheets via a combination of traditional vision techniques and state-of-the-art deep learning models to identify and isolate pipeline codes, pipelines, inlets and outlets, and for detecting symbols. This is followed by association of the detected components with the appropriate pipeline. The extracted pipeline information is used to populate a tree-like data-structure for capturing the structure of the piping schematics. We evaluated proposed method on a real world dataset of P&ID sheets obtained from an oil firm and have obtained promising results. | There exists very limited work on digitizing the content of engineering diagrams to facilitate fast and efficient extraction of information. The authors @cite_14 automated the assessment of AutoCAD Drawing Exchange Format (DXF) by converting DXF file into SVG format and developing a marking algorithm of the generated SVG files. A framework for engineering drawings recognition using a case-based approach is proposed by @cite_10 where the user interactively provides an example of one type of graphic object in an engineering drawing and then system tries to learn the graphical knowledge of this type of graphic object from the example and later use this learned knowledge to recognize or search for similar graphic objects in engineering drawings. Authors of @cite_6 tried to automate the extraction of structural and connectivity information from vector-graphics-coded engineering documents. A spatial relation graph (SRG) and its partial matching method are proposed for online composite graphics representation and recognition in @cite_3 . Overall, we observed that there does not exist much work on information extraction from plant engineering diagrams. | {
"cite_N": [
"@cite_3",
"@cite_14",
"@cite_10",
"@cite_6"
],
"mid": [
"1583110160",
"2107266793",
"2210190747",
"2031948510"
],
"abstract": [
"In this paper, we propose a framework for engineering drawings recognition using a case-based approach. The key idea of our scheme is that, interactively, the user provides an example of one type of graphic object in an engineering drawing, then the system learns the graphical knowledge of this type of graphic object from the example and uses this learned knowledge to recognize or search for similar graphic objects in engineering drawings. The scheme emphasizes the following three distinct characteristics: automatism, run-time-ness, and robustness. We summarized five types of geometric constraints to represent the generic graphical knowledge. We also developed two algorithms for case-based graphical knowledge acquisition and knowledge-based graphics recognition, respectively. Experiments have shown that our proposed framework is both efficient and effective for recognizing various types of graphic objects in engineering drawings.",
"A spatial relation graph (SRG) and its partial matching method are proposed for online composite graphics representation and recognition. The SRG-based approach emphasizes three characteristics of online graphics recognition: partial, structural, and independent of stroke order and stroke number. A constrained partial permutation strategy is also proposed to reduce the computational cost of matching two SRGs, which is originally an NP-complete problem as is graph isomorphism. Experimental results show that our proposed SRG-based approach is both efficient and effective for online composite graphics recognition in our sketch-based graphics input system - SmartSketchpad.",
"Assessment of student's Engineering Drawing (ED) is always tedious, repetitive and time consuming. Image processing has been the common method to convert ED to be automatically assessed. This method is tedious as algorithms need to be developed for each shape to be assessed. Our research aims to create a software application that is able to perform automatic assessment for AutoCAD Drawing Exchange Format (DXF) files for undergraduate ED course. To achieve this goal, we have explored methods to convert DXF files into SVG format and develop a marking algorithm for the generated SVG files. The result shows that it is feasible to create software that automatically assesses ED without human intervention. Future implementation would include complex real-world ED.",
"We present our recent model of a diagram recognition engine. It extends our previous work which approaches the structural recognition as an optimization problem of choosing the best subset of symbol candidates. The main improvement is the integration of our own text separator into the pipeline to deal with text blocks occurring in diagrams. Second improvement is splitting the symbol candidates detection into two stages: uniform symbols detection and arrows detection. Text recognition is left for post processing when the diagram structure is already known. Training and testing of the engine was done on a freely available benchmark database of flowcharts. We correctly segmented and recognized 93.0 of the symbols having 55.1 of the diagrams recognized without any error. Considering correct stroke labeling, we achieved the precision of 95.7 . This result is superior to the state-of-the-art method with the precision of 92.4 . Additionally, we demonstrate the generality of the proposed method by adapting the system to finite automata domain and evaluating it on own database of such diagrams."
]
} |
1901.11383 | 2952973608 | One of the most common modes of representing engineering schematics are Piping and Instrumentation diagrams (P&IDs) that describe the layout of an engineering process flow along with the interconnected process equipment. Over the years, P&ID diagrams have been manually generated, scanned and stored as image files. These files need to be digitized for purposes of inventory management and updation, and easy reference to different components of the schematics. There are several challenging vision problems associated with digitizing real world P&ID diagrams. Real world P&IDs come in several different resolutions, and often contain noisy textual information. Extraction of instrumentation information from these diagrams involves accurate detection of symbols that frequently have minute visual differences between them. Identification of pipelines that may converge and diverge at different points in the image is a further cause for concern. Due to these reasons, to the best of our knowledge, no system has been proposed for end-to-end data extraction from P&ID diagrams. However, with the advent of deep learning and the spectacular successes it has achieved in vision, we hypothesized that it is now possible to re-examine this problem armed with the latest deep learning models. To that end, we present a novel pipeline for information extraction from P&ID sheets via a combination of traditional vision techniques and state-of-the-art deep learning models to identify and isolate pipeline codes, pipelines, inlets and outlets, and for detecting symbols. This is followed by association of the detected components with the appropriate pipeline. The extracted pipeline information is used to populate a tree-like data-structure for capturing the structure of the piping schematics. We evaluated proposed method on a real world dataset of P&ID sheets obtained from an oil firm and have obtained promising results. | However, we discovered a significant body of work on recognition of symbols in prior art. @cite_19 proposed Fourier Mellin Transform features to classify multi-oriented and multi-scaled patterns in engineering diagrams. Other models utilized for symbol recognition include Auto Associative neural networks @cite_4 , Deep Belief networks @cite_20 , and consistent attributed graphs (CAG) @cite_15 . There are also models that use a set of visual features which capture online stroke properties like orientation and endpoint location @cite_16 , and shape based matching between different symbols @cite_11 . We see that most of the prior work focuses on extracting symbols from such engineering diagrams or flow charts. To the best of our knowledge, there exists no work which has proposed an end-to-end pipeline for automating the information extraction from plant engineering diagrams such as P &ID. | {
"cite_N": [
"@cite_4",
"@cite_19",
"@cite_15",
"@cite_16",
"@cite_20",
"@cite_11"
],
"mid": [
"2031948510",
"2798365843",
"1929903369",
"2125380992",
"2729172879",
"2527800741"
],
"abstract": [
"We present our recent model of a diagram recognition engine. It extends our previous work which approaches the structural recognition as an optimization problem of choosing the best subset of symbol candidates. The main improvement is the integration of our own text separator into the pipeline to deal with text blocks occurring in diagrams. Second improvement is splitting the symbol candidates detection into two stages: uniform symbols detection and arrows detection. Text recognition is left for post processing when the diagram structure is already known. Training and testing of the engine was done on a freely available benchmark database of flowcharts. We correctly segmented and recognized 93.0 of the symbols having 55.1 of the diagrams recognized without any error. Considering correct stroke labeling, we achieved the precision of 95.7 . This result is superior to the state-of-the-art method with the precision of 92.4 . Additionally, we demonstrate the generality of the proposed method by adapting the system to finite automata domain and evaluating it on own database of such diagrams.",
"Compared to earlier multistage frameworks using CNN features, recent end-to-end deep approaches for fine-grained recognition essentially enhance the mid-level learning capability of CNNs. Previous approaches achieve this by introducing an auxiliary network to infuse localization information into the main classification network, or a sophisticated feature encoding method to capture higher order feature statistics. We show that mid-level representation learning can be enhanced within the CNN framework, by learning a bank of convolutional filters that capture class-specific discriminative patches without extra part or bounding box annotations. Such a filter bank is well structured, properly initialized and discriminatively learned through a novel asymmetric multi-stream architecture with convolutional filter supervision and a non-random layer initialization. Experimental results show that our approach achieves state-of-the-art on three publicly available fine-grained recognition datasets (CUB-200-2011, Stanford Cars and FGVC-Aircraft). Ablation studies and visualizations are provided to understand our approach.",
"Deep convolutional neural networks (CNN) have seen tremendous success in large-scale generic object recognition. In comparison with generic object recognition, fine-grained image classification (FGIC) is much more challenging because (i) fine-grained labeled data is much more expensive to acquire (usually requiring domain expertise); (ii) there exists large intra-class and small inter-class variance. Most recent work exploiting deep CNN for image recognition with small training data adopts a simple strategy: pre-train a deep CNN on a large-scale external dataset (e.g., ImageNet) and fine-tune on the small-scale target data to fit the specific classification task. In this paper, beyond the fine-tuning strategy, we propose a systematic framework of learning a deep CNN that addresses the challenges from two new perspectives: (i) identifying easily annotated hyper-classes inherent in the fine-grained data and acquiring a large number of hyper-class-labeled images from readily available external sources (e.g., image search engines), and formulating the problem into multitask learning; (ii) a novel learning model by exploiting a regularization between the fine-grained recognition model and the hyper-class recognition model. We demonstrate the success of the proposed framework on two small-scale fine-grained datasets (Stanford Dogs and Stanford Cars) and on a large-scale car dataset that we collected.",
"Symbol recognition is a well-known problem in the field of graphics. A symbol can be defined as a structure within document that has a particular meaning in the context of the application. Due to their representational power, graph structures are usually used to represent line drawings images.An accurate vectorization constitutes a first approach to solve this goal. But vectorization only gives the segments constituting the document and their geometrical attributes.Interpreting a document such as P&ID (Process & Instrumentation)diagram requires an additional stage viz. recognition of symbols in terms of its shape. Usually a P&ID diagram contain several types of elements, symbols and structural connectivity. For those symbols that can be defined by a prototype pattern, we propose an iterative learning strategy based on Hopfield model to learn the symbols, for subsequent recognition in the P&ID diagram. In a typical shape recognition problem one has to account for transformation invariance. Here the transformation invariance is circumvented by using an iterative learning approach which can learn symbols with high degree of correlation.",
"We present a simple and effective architecture for fine-grained visual recognition called Bilinear Convolutional Neural Networks (B-CNNs). These networks represent an image as a pooled outer product of features derived from two CNNs and capture localized feature interactions in a translationally invariant manner. B-CNNs belong to the class of orderless texture representations but unlike prior work they can be trained in an end-to-end manner. Our most accurate model obtains 84.1 , 79.4 , 86.9 and 91.3 per-image accuracy on the Caltech-UCSD birds [67], NABirds [64], FGVC aircraft [42], and Stanford cars [33] dataset respectively and runs at 30 frames-per-second on a NVIDIA Titan X GPU. We then present a systematic analysis of these networks and show that (1) the bilinear features are highly redundant and can be reduced by an order of magnitude in size without significant loss in accuracy, (2) are also effective for other image classification tasks such as texture and scene recognition, and (3) can be trained from scratch on the ImageNet dataset offering consistent improvements over the baseline architecture. Finally, we present visualizations of these models on various datasets using top activations of neural units and gradient-based inversion techniques. The source code for the complete system is available at this http URL.",
"Convolutional neural networks (CNNs) have made remarkable progress on scene recognition, partially due to these recent large-scale scene datasets, such as the Places and Places2. Scene categories are often defined by multi-level information, including local objects, global layout, and background environment, thus leading to large intra-class variations. In addition, with the increasing number of scene categories, label ambiguity has become another crucial issue in large-scale classification. This paper focuses on large-scale scene recognition and makes two major contributions to tackle these issues. First, we propose a multi-resolution CNN architecture that captures visual content and structure at multiple levels. The multi-resolution CNNs are composed of coarse resolution CNNs and fine resolution CNNs, which are complementary to each other. Second, we design two knowledge guided disambiguation techniques to deal with the problem of label ambiguity: 1) we exploit the knowledge from the confusion matrix computed on validation data to merge ambiguous classes into a super category and 2) we utilize the knowledge of extra networks to produce a soft label for each image. Then, the super categories or soft labels are employed to guide CNN training on the Places2. We conduct extensive experiments on three large-scale image datasets (ImageNet, Places, and Places2), demonstrating the effectiveness of our approach. Furthermore, our method takes part in two major scene recognition challenges, and achieves the second place at the Places2 challenge in ILSVRC 2015, and the first place at the LSUN challenge in CVPR 2016. Finally, we directly test the learned representations on other scene benchmarks, and obtain the new state-of-the-art results on the MIT Indoor67 (86.7 ) and SUN397 (72.0 ). We release the code and models at https: github.com wanglimin MRCNN-Scene-Recognition ."
]
} |
1901.11383 | 2952973608 | One of the most common modes of representing engineering schematics are Piping and Instrumentation diagrams (P&IDs) that describe the layout of an engineering process flow along with the interconnected process equipment. Over the years, P&ID diagrams have been manually generated, scanned and stored as image files. These files need to be digitized for purposes of inventory management and updation, and easy reference to different components of the schematics. There are several challenging vision problems associated with digitizing real world P&ID diagrams. Real world P&IDs come in several different resolutions, and often contain noisy textual information. Extraction of instrumentation information from these diagrams involves accurate detection of symbols that frequently have minute visual differences between them. Identification of pipelines that may converge and diverge at different points in the image is a further cause for concern. Due to these reasons, to the best of our knowledge, no system has been proposed for end-to-end data extraction from P&ID diagrams. However, with the advent of deep learning and the spectacular successes it has achieved in vision, we hypothesized that it is now possible to re-examine this problem armed with the latest deep learning models. To that end, we present a novel pipeline for information extraction from P&ID sheets via a combination of traditional vision techniques and state-of-the-art deep learning models to identify and isolate pipeline codes, pipelines, inlets and outlets, and for detecting symbols. This is followed by association of the detected components with the appropriate pipeline. The extracted pipeline information is used to populate a tree-like data-structure for capturing the structure of the piping schematics. We evaluated proposed method on a real world dataset of P&ID sheets obtained from an oil firm and have obtained promising results. | In literature, Connected Component (CC) analysis @cite_18 has been used extensively for extracting characters @cite_13 from images. However, connected components are extremely sensitive to noise and thresholding may not be suitable for P &ID text extraction. Hence, we utilize the recently invented Connectionist Temporal Proposal Network (CTPN) @cite_0 to detect text in the image with impressive accuracy. For line detection, we utilize Probabilistic hough transform (PHT) @cite_7 which is computationally efficient and fast version of the standard hough transform as it uses random sampling of edge points to find lines present in the image. We make use of PHT for determining all the lines present in P &ID sheets which are possible candidates for pipelines. In our paper, we propose the use of Fully convolutional neural network (FCN) based segmentation @cite_1 for detecting symbols because tranditional classification networks were unable to differentiate among different types of symbols due to very minute inter-class differences in visual appearances and presence of noisy and textual information present inside symbols. FCN incorporates contextual as well as spatial relationship of symbols in the image, which is often necessary for accurate detection and classification of P &ID symbols. | {
"cite_N": [
"@cite_18",
"@cite_7",
"@cite_1",
"@cite_0",
"@cite_13"
],
"mid": [
"2743620784",
"2963727650",
"2519818067",
"2789983685",
"2472159136"
],
"abstract": [
"The region-based Convolutional Neural Network (CNN) detectors such as Faster R-CNN or R-FCN have already shown promising results for object detection by combining the region proposal subnetwork and the classification subnetwork together. Although R-FCN has achieved higher detection speed while keeping the detection performance, the global structure information is ignored by the position-sensitive score maps. To fully explore the local and global properties, in this paper, we propose a novel fully convolutional network, named as CoupleNet, to couple the global structure with local parts for object detection. Specifically, the object proposals obtained by the Region Proposal Network (RPN) are fed into the the coupling module which consists of two branches. One branch adopts the position-sensitive RoI (PSRoI) pooling to capture the local part information of the object, while the other employs the RoI pooling to encode the global and context information. Next, we design different coupling strategies and normalization ways to make full use of the complementary advantages between the global and local branches. Extensive experiments demonstrate the effectiveness of our approach. We achieve state-of-the-art results on all three challenging datasets, i.e. a mAP of 82.7 on VOC07, 80.4 on VOC12, and 34.4 on COCO. Codes will be made publicly available.",
"Recent work has made significant progress in improving spatial resolution for pixelwise labeling with Fully Convolutional Network (FCN) framework by employing Dilated Atrous convolution, utilizing multi-scale features and refining boundaries. In this paper, we explore the impact of global contextual information in semantic segmentation by introducing the Context Encoding Module, which captures the semantic context of scenes and selectively highlights class-dependent featuremaps. The proposed Context Encoding Module significantly improves semantic segmentation results with only marginal extra computation cost over FCN. Our approach has achieved new state-of-the-art results 51.7 mIoU on PASCAL-Context, 85.9 mIoU on PASCAL VOC 2012. Our single model achieves a final score of 0.5567 on ADE20K test set, which surpasses the winning entry of COCO-Place Challenge 2017. In addition, we also explore how the Context Encoding Module can improve the feature representation of relatively shallow networks for the image classification on CIFAR-10 dataset. Our 14 layer network has achieved an error rate of 3.45 , which is comparable with state-of-the-art approaches with over 10A— more layers. The source code for the complete system are publicly available1.",
"We propose a novel Connectionist Text Proposal Network (CTPN) that accurately localizes text lines in natural image. The CTPN detects a text line in a sequence of fine-scale text proposals directly in convolutional feature maps. We develop a vertical anchor mechanism that jointly predicts location and text non-text score of each fixed-width proposal, considerably improving localization accuracy. The sequential proposals are naturally connected by a recurrent neural network, which is seamlessly incorporated into the convolutional network, resulting in an end-to-end trainable model. This allows the CTPN to explore rich context information of image, making it powerful to detect extremely ambiguous text. The CTPN works reliably on multi-scale and multi-language text without further post-processing, departing from previous bottom-up methods requiring multi-step post filtering. It achieves 0.88 and 0.61 F-measure on the ICDAR 2013 and 2015 benchmarks, surpassing recent results [8, 35] by a large margin. The CTPN is computationally efficient with 0.14 s image, by using the very deep VGG16 model [27]. Online demo is available: http: textdet.com .",
"Recent work has made significant progress in improving spatial resolution for pixelwise labeling with Fully Convolutional Network (FCN) framework by employing Dilated Atrous convolution, utilizing multi-scale features and refining boundaries. In this paper, we explore the impact of global contextual information in semantic segmentation by introducing the Context Encoding Module, which captures the semantic context of scenes and selectively highlights class-dependent featuremaps. The proposed Context Encoding Module significantly improves semantic segmentation results with only marginal extra computation cost over FCN. Our approach has achieved new state-of-the-art results 51.7 mIoU on PASCAL-Context, 85.9 mIoU on PASCAL VOC 2012. Our single model achieves a final score of 0.5567 on ADE20K test set, which surpass the winning entry of COCO-Place Challenge in 2017. In addition, we also explore how the Context Encoding Module can improve the feature representation of relatively shallow networks for the image classification on CIFAR-10 dataset. Our 14 layer network has achieved an error rate of 3.45 , which is comparable with state-of-the-art approaches with over 10 times more layers. The source code for the complete system are publicly available.",
"We propose a system that finds text in natural scenes using a variety of cues. Our novel data-driven method incorporates coarse-to-fine detection of character pixels using convolutional features (Text-Conv), followed by extracting connected components (CCs) from characters using edge and color features, and finally performing a graph-based segmentation of CCs into words (Word-Graph). For Text-Conv, the initial detection is based on convolutional feature maps similar to those used in Convolutional Neural Networks (CNNs), but learned using Convolutional k-means. Convolution masks defined by local and neighboring patch features are used to improve detection accuracy. The Word-Graph algorithm uses contextual information to both improve word segmentation and prune false character word detections. Different definitions for foreground (text) regions are used to train the detection stages, some based on bounding box intersection, and others on bounding box and pixel intersection. Our system obtains pixel, character, and word detection f-measures of 93.14 , 90.26 , and 86.77 respectively for the ICDAR 2015 Robust Reading Focused Scene Text dataset, out-performing state-of-the-art systems. This approach may work for other detection targets with homogenous color in natural scenes."
]
} |
1901.11417 | 2913832161 | Fluid approximations have seen great success in approximating the macro-scale behaviour of Markov systems with a large number of discrete states. However, these methods rely on the continuous-time Markov chain (CTMC) having a particular population structure which suggests a natural continuous state-space endowed with a dynamics for the approximating process. We construct here a general method based on spectral analysis of the transition matrix of the CTMC, without the need for a population structure. Specifically, we use the popular manifold learning method of diffusion maps to analyse the transition matrix as the operator of a hidden continuous process. An embedding of states in a continuous space is recovered, and the space is endowed with a drift vector field inferred via Gaussian process regression. In this manner, we construct an ODE whose solution approximates the evolution of the CTMC mean, mapped onto the continuous space (known as the fluid limit). | In the case of pCTMCs, a more concise description in terms of the collective dynamics of population averages is however available. Starting with the seminal work of van Kampen @cite_0 , and motivated by the interpretation of pCTMCs as chemical reaction systems, several approximation schemes have been developed which relax the original pCTMC to a continuous stochastic process; see @cite_8 for a recent review. | {
"cite_N": [
"@cite_0",
"@cite_8"
],
"mid": [
"2150055571",
"2046451925"
],
"abstract": [
"In the stochastic formulation of chemical reactions, the dynamics of the first M-order moments of the species populations generally do not form a closed system of differential equations, in the sense that the time-derivatives of first M-order moments generally depend on moments of order higher than M. However, for analysis purposes, these dynamics are often made to be closed by approximating the needed derivatives of the first M-order moments by nonlinear functions of the same moments. These functions are called the moment closure functions. Recent results have introduced the technique of derivative-matching, where the moment closure functions are obtained by first assuming that they exhibit a certain separable form, and then matching time derivatives of the exact (not closed) moment equations with that of the approximate (closed) equations for some initial time and set of initial conditions. However, for multi-species reactions these results have been restricted to second order truncations, i.e, M = 2. This paper extends these results by providing explicit formulas to construct moment closure functions for any arbitrary order of truncation M. We show that with increasing M the closed moment equations provide more accurate approximations to the exact moment equations. Striking features of these moment closure functions are that they are independent of the reaction parameters (reaction rates and stoichiometry) and moreover the dependence of higher-order moment on lower order ones is consistent with the population being jointly lognormally distributed. To illustrate the applicability of our results we consider a simple bi-molecular reaction. Moment estimates from a third order truncation are compared with estimates obtained from a large number of Monte Carlo simulations",
"The stochastic dynamical behavior of a well-stirred mixture of N molecular species that chemically interact through M reaction channels is accurately described by the chemical master equation. It is shown here that, whenever two explicit dynamical conditions are satisfied, the microphysical premise from which the chemical master equation is derived leads directly to an approximate time-evolution equation of the Langevin type. This chemical Langevin equation is the same as one studied earlier by Kurtz, in contradistinction to some other earlier proposed forms that assume a deterministic macroscopic evolution law. The novel aspect of the present analysis is that it shows that the accuracy of the equation depends on the satisfaction of certain specific conditions that can change from moment to moment, rather than on a static system size parameter. The derivation affords a new perspective on the origin and magnitude of noise in a chemically reacting system. It also clarifies the connection between the stochas..."
]
} |
1901.11417 | 2913832161 | Fluid approximations have seen great success in approximating the macro-scale behaviour of Markov systems with a large number of discrete states. However, these methods rely on the continuous-time Markov chain (CTMC) having a particular population structure which suggests a natural continuous state-space endowed with a dynamics for the approximating process. We construct here a general method based on spectral analysis of the transition matrix of the CTMC, without the need for a population structure. Specifically, we use the popular manifold learning method of diffusion maps to analyse the transition matrix as the operator of a hidden continuous process. An embedding of states in a continuous space is recovered, and the space is endowed with a drift vector field inferred via Gaussian process regression. In this manner, we construct an ODE whose solution approximates the evolution of the CTMC mean, mapped onto the continuous space (known as the fluid limit). | Following Darling and Norris @cite_17 , we examine and formalise the aspects of pCTMCs which render them especially amenable to the fluid approximation. As mentioned, the first is that pCTMC state-spaces are countable and there exists an obvious ordering. We can therefore write a trivial linear mapping from the discrete, countable state-space @math to a continuous Euclidean space @math , where @math is the number of agent types in the system. | {
"cite_N": [
"@cite_17"
],
"mid": [
"2131441094"
],
"abstract": [
"This paper concerns computational methods for verifying properties of polyhedral invariant hybrid automata (PIHA), which are hybrid automata with discrete transitions governed by polyhedral guards. To verify properties of the state trajectories for PIHA, the planar switching surfaces are partitioned to define a finite set of discrete states in an approximate quotient transition system (AQTS). State transitions in the AQTS are determined by the reachable states, or flow pipes, emitting from the switching surfaces according to the continuous dynamics. This paper presents a method for computing polyhedral approximations to flow pipes. It is shown that the flow-pipe approximation error can be made arbitrarily small for general nonlinear dynamics and that the computations can be made more efficient for affine systems. The paper also describes CheckMate, a MATLAB-based tool for modeling, simulating and verifying properties of hybrid systems based on the computational methods previously described."
]
} |
1901.11417 | 2913832161 | Fluid approximations have seen great success in approximating the macro-scale behaviour of Markov systems with a large number of discrete states. However, these methods rely on the continuous-time Markov chain (CTMC) having a particular population structure which suggests a natural continuous state-space endowed with a dynamics for the approximating process. We construct here a general method based on spectral analysis of the transition matrix of the CTMC, without the need for a population structure. Specifically, we use the popular manifold learning method of diffusion maps to analyse the transition matrix as the operator of a hidden continuous process. An embedding of states in a continuous space is recovered, and the space is endowed with a drift vector field inferred via Gaussian process regression. In this manner, we construct an ODE whose solution approximates the evolution of the CTMC mean, mapped onto the continuous space (known as the fluid limit). | There are many ways to satisfy the above criteria, but a common one (used in pCTMCs) is hydrodynamic scaling'', where the increments of the @math -state Markov process mapped to the Euclidean space are @math and the jump rate is @math . The criteria above are derived formally in @cite_17 @cite_18 and ensure that: | {
"cite_N": [
"@cite_18",
"@cite_17"
],
"mid": [
"2046434141",
"143740007"
],
"abstract": [
"Consider a graph, G, for which the vertices can have two modes, 0 or 1. Suppose that a particle moves around on G according to a discrete time Markov chain with the following rules. With (strictly positive) probabilities pm, pc and pr it moves to a randomly chosen neighbour, changes the mode of the vertex it is at or just stands still, respectively. We call such a random process a (pm, pc, pr)-lamplighter process on G. Assume that the process starts with the particle in a fixed position and with all vertices having mode 0. The convergence rate to stationarity in terms of the total variation norm is studied for the special cases with G = KN, the complete graph with N vertices, and when G = mod N. In the former case we prove that as N --> [infinity], ((2pc + pm) 4pcpm)N log N is a threshold for the convergence rate. In the latter case we show that the convergence rate is asymptotically determined by the cover time CN in that the total variation norm after aN2 steps is given by P(CN > aN2). The limit of this probability can in turn be calculated by considering a Brownian motion with two absorbing barriers. In particular, this means that there is no threshold for this case.",
"From a theoretical perspective, scale invariance, or simply scaling, can fruitfully be modeled with classes of multifractal stochastic processes, designed from positive multiplicative martingales (or cascades). From a practical perspective, scaling in real-world data is often analyzed by means of multiresolution quantities. The present contribution focuses on three different types of such multiresolution quantities, namely increment, wavelet and Leader coefficients, as well as on a specific multifractal processes, referred to as Infinitely Divisible Motions and fractional Brownian motion in multifractal time. It aims at studying, both analytically and by numerical simulations, the impact of varying the number of vanishing moments of the mother wavelet and the order of the increments on the decay rate of the (higher order) covariance functions of the (q-th power of the absolute values of these) multiresolution coefficients. The key result obtained here consist of the fact that, though it fastens the decay of the covariance functions, as is the case for fractional Brownian motions, increasing the number of vanishing moments of the mother wavelet or the order of the increments does not induce any faster decay for the (higher order) covariance functions"
]
} |
1901.11344 | 2914116739 | Recent years has witnessed dramatic progress of neural machine translation (NMT), however, the method of manually guiding the translation procedure remains to be better explored. Previous works proposed to handle such problem through lexcially-constrained beam search in the decoding phase. Unfortunately, these lexically-constrained beam search methods suffer two fatal disadvantages: high computational complexity and hard beam search which generates unexpected translations. In this paper, we propose to learn the ability of lexically-constrained translation with external memory, which can overcome the above mentioned disadvantages. For the training process, automatically extracted phrase pairs are extracted from alignment and sentence parsing, then further be encoded into an external memory. This memory is then used to provide lexically-constrained information for training through a memory-attention machanism. Various experiments are conducted on WMT Chinese to English and English to German tasks. All the results can demonstrate the effectiveness of our method. | The establishment of one efficient and effective machine translation system is attractive over the decades. Although systems based on statistical machine translation have been used in real life, the unpromising performance makes it difficult to be promoted. Recent works of neural machine translation have made this possible. proposed an attention mechanism for encoder-decoder neural machine translation system, which can sufficiently explore the context representation in the source sentences. Transformer @cite_1 is a more promising neural machine translation architecture with self-attention, which can achieve faster training speed and better performance. | {
"cite_N": [
"@cite_1"
],
"mid": [
"2909620036"
],
"abstract": [
"Neural machine translation has significantly pushed forward the quality of the field. However, there are remaining big issues with the translations and one of them is fairness. Neural models are trained on large text corpora which contains biases and stereotypes. As a consequence, models inherit these social biases. Recent methods have shown results in reducing gender bias in other natural language processing applications such as word embeddings. We take advantage of the fact that word embeddings are used in neural machine translation to propose the first debiased machine translation system. Specifically, we propose, experiment and analyze the integration of two debiasing techniques over GloVe embeddings in the Transformer translation architecture. We evaluate our proposed system on a generic English-Spanish task, showing gains up to one BLEU point. As for the gender bias evaluation, we generate a test set of occupations and we show that our proposed system learns to equalize existing biases from the baseline system."
]
} |
1901.11344 | 2914116739 | Recent years has witnessed dramatic progress of neural machine translation (NMT), however, the method of manually guiding the translation procedure remains to be better explored. Previous works proposed to handle such problem through lexcially-constrained beam search in the decoding phase. Unfortunately, these lexically-constrained beam search methods suffer two fatal disadvantages: high computational complexity and hard beam search which generates unexpected translations. In this paper, we propose to learn the ability of lexically-constrained translation with external memory, which can overcome the above mentioned disadvantages. For the training process, automatically extracted phrase pairs are extracted from alignment and sentence parsing, then further be encoded into an external memory. This memory is then used to provide lexically-constrained information for training through a memory-attention machanism. Various experiments are conducted on WMT Chinese to English and English to German tasks. All the results can demonstrate the effectiveness of our method. | External memory has been used in several works @cite_2 @cite_0 @cite_3 to enhance the quality of neural machine translation. For example, proposes to extract phrase table as recommendation memory for neural machine translation. However, this kind of phrase table is too noisy, which is also mentioned in . proposes to store the hidden context information into the memory, which can be used to calculate an additional probability of target word. Both of these two methods require a high quality translation alignment. @cite_4 proposes to annotate the source sentences with experts and use a copy-generator for rare word translation. However, the strong copy ability may cuase the loss of fluency. And aims to improve the performance of NMT by maintaining a updatable memory. | {
"cite_N": [
"@cite_0",
"@cite_4",
"@cite_3",
"@cite_2"
],
"mid": [
"2951132175",
"2795933031",
"2769311391",
"2618463334"
],
"abstract": [
"Neural Machine Translation (NMT) has drawn much attention due to its promising translation performance recently. However, several studies indicate that NMT often generates fluent but unfaithful translations. In this paper, we propose a method to alleviate this problem by using a phrase table as recommendation memory. The main idea is to add bonus to words worthy of recommendation, so that NMT can make correct predictions. Specifically, we first derive a prefix tree to accommodate all the candidate target phrases by searching the phrase translation table according to the source sentence. Then, we construct a recommendation word set by matching between candidate target phrases and previously translated target words by NMT. After that, we determine the specific bonus value for each recommendable word by using the attention vector and phrase translation probability. Finally, we integrate this bonus value into NMT to improve the translation results. The extensive experiments demonstrate that the proposed methods obtain remarkable improvements over the strong attentionbased NMT.",
"One of the difficulties of neural machine translation (NMT) is the recall and appropriate translation of low-frequency words or phrases. In this paper, we propose a simple, fast, and effective method for recalling previously seen translation examples and incorporating them into the NMT decoding process. Specifically, for an input sentence, we use a search engine to retrieve sentence pairs whose source sides are similar with the input sentence, and then collect @math -grams that are both in the retrieved target sentences and aligned with words that match in the source sentences, which we call \"translation pieces\". We compute pseudo-probabilities for each retrieved sentence based on similarities between the input sentence and the retrieved source sentences, and use these to weight the retrieved translation pieces. Finally, an existing NMT model is used to translate the input sentence, with an additional bonus given to outputs that contain the collected translation pieces. We show our method improves NMT translation results up to 6 BLEU points on three narrow domain translation tasks where repetitiveness of the target sentences is particularly salient. It also causes little increase in the translation time, and compares favorably to another alternative retrieval-based method with respect to accuracy, speed, and simplicity of implementation.",
"Existing neural machine translation (NMT) models generally translate sentences in isolation, missing the opportunity to take advantage of document-level information. In this work, we propose to augment NMT models with a very light-weight cache-like memory network, which stores recent hidden representations as translation history. The probability distribution over generated words is updated online depending on the translation history retrieved from the memory, endowing NMT models with the capability to dynamically adapt over time. Experiments on multiple domains with different topics and styles show the effectiveness of the proposed approach with negligible impact on the computational cost.",
"In this paper, we extend an attention-based neural machine translation (NMT) model by allowing it to access an entire training set of parallel sentence pairs even after training. The proposed approach consists of two stages. In the first stage--retrieval stage--, an off-the-shelf, black-box search engine is used to retrieve a small subset of sentence pairs from a training set given a source sentence. These pairs are further filtered based on a fuzzy matching score based on edit distance. In the second stage--translation stage--, a novel translation model, called translation memory enhanced NMT (TM-NMT), seamlessly uses both the source sentence and a set of retrieved sentence pairs to perform the translation. Empirical evaluation on three language pairs (En-Fr, En-De, and En-Es) shows that the proposed approach significantly outperforms the baseline approach and the improvement is more significant when more relevant sentence pairs were retrieved."
]
} |
1901.11259 | 2913092861 | Convolutional neural networks have been widely used in content-based image retrieval. To better deal with large-scale data, the deep hashing model is proposed as an effective method, which maps an image to a binary code that can be used for hashing search. However, most existing deep hashing models only utilize fine-level semantic labels or convert them to similar dissimilar labels for training. The natural semantic hierarchy structures are ignored in the training stage of the deep hashing model. In this paper, we present an effective algorithm to train a deep hashing model that can preserve a semantic hierarchy structure for large-scale image retrieval. Experiments on two datasets show that our method improves the fine-level retrieval performance. Meanwhile, our model achieves state-of-the-art results in terms of hierarchical retrieval. | DSPH @cite_13 firstly proposes to utilize the pair-wise label to train the end-to-end deep hashing model. HashNet @cite_0 defines a weighted maximum likelihood of pairwise logistic for balance the similar and dissimilar labels. DTSH @cite_22 extends the pair-wise supervision to the triplet one for more effectively capturing the semantic information. To fully use the semantic class labels, several works design loss functions directly based on the class labels. SSDH @cite_17 utilizes the softmax classifier to train the hashing model. DCWH @cite_12 constructs a Gaussian distribution-based objective function to take the advantage of the class-level information. DSRH @cite_15 and DSDH @cite_10 both combine the pair-wise and class-level supervisions. | {
"cite_N": [
"@cite_22",
"@cite_10",
"@cite_0",
"@cite_15",
"@cite_13",
"@cite_12",
"@cite_17"
],
"mid": [
"1956333070",
"2791396492",
"2531549126",
"2586937979",
"1996219872",
"2963240808",
"2773012335"
],
"abstract": [
"In this paper, we propose a new deep hashing (DH) approach to learn compact binary codes for large scale visual search. Unlike most existing binary codes learning methods which seek a single linear projection to map each sample into a binary vector, we develop a deep neural network to seek multiple hierarchical non-linear transformations to learn these binary codes, so that the nonlinear relationship of samples can be well exploited. Our model is learned under three constraints at the top layer of the deep network: 1) the loss between the original real-valued feature descriptor and the learned binary vector is minimized, 2) the binary codes distribute evenly on each bit, and 3) different bits are as independent as possible. To further improve the discriminative power of the learned binary codes, we extend DH into supervised DH (SDH) by including one discriminative term into the objective function of DH which simultaneously maximizes the inter-class variations and minimizes the intra-class variations of the learned binary codes. Experimental results show the superiority of the proposed approach over the state-of-the-arts.",
"Deep supervised hashing has emerged as an influential solution to large-scale semantic image retrieval problems in computer vision. In the light of recent progress, convolutional neural network based hashing methods typically seek pair-wise or triplet labels to conduct the similarity preserving learning. However, complex semantic concepts of visual contents are hard to capture by similar dissimilar labels, which limits the retrieval performance. Generally, pair-wise or triplet losses not only suffer from expensive training costs but also lack in extracting sufficient semantic information. In this regard, we propose a novel deep supervised hashing model to learn more compact class-level similarity preserving binary codes. Our deep learning based model is motivated by deep metric learning that directly takes semantic labels as supervised information in training and generates corresponding discriminant hashing code. Specifically, a novel cubic constraint loss function based on Gaussian distribution is proposed, which preserves semantic variations while penalizes the overlap part of different classes in the embedding space. To address the discrete optimization problem introduced by binary codes, a two-step optimization strategy is proposed to provide efficient training and avoid the problem of gradient vanishing. Extensive experiments on four large-scale benchmark databases show that our model can achieve the state-of-the-art retrieval performance. Moreover, when training samples are limited, our method surpasses other supervised deep hashing methods with non-negligible margins.",
"Hashing techniques have been intensively investigated for large scale vision applications. Recent research has shown that leveraging supervised information can lead to high quality hashing. However, most existing supervised hashing methods only construct similarity-preserving hash codes. Observing that semantic structures carry complementary information, we propose the idea of cotraining for hashing, by jointly learning projections from image representations to hash codes and classification. Specifically, a novel deep semantic-preserving and ranking-based hashing (DSRH) architecture is presented, which consists of three components: a deep CNN for learning image representations, a hash stream of a binary mapping layer by evenly dividing the learnt representations into multiple bags and encoding each bag into one hash bit, and a classification stream. Mean-while, our model is learnt under two constraints at the top loss layer of hash stream: a triplet ranking loss and orthogonality constraint. The former aims to preserve the relative similarity ordering in the triplets, while the latter makes different hash bit as independent as possible. We have conducted experiments on CIFAR-10 and NUS-WIDE image benchmarks, demonstrating that our approach can provide superior image search accuracy than other state-of-the-art hashing techniques.",
"This paper presents a simple yet effective supervised deep hash approach that constructs binary hash codes from labeled data for large-scale image search. We assume that the semantic labels are governed by several latent attributes with each attribute on or off , and classification relies on these attributes. Based on this assumption, our approach, dubbed supervised semantics-preserving deep hashing (SSDH), constructs hash functions as a latent layer in a deep network and the binary codes are learned by minimizing an objective function defined over classification error and other desirable hash codes properties. With this design, SSDH has a nice characteristic that classification and retrieval are unified in a single learning model. Moreover, SSDH performs joint learning of image representations, hash codes, and classification in a point-wised manner, and thus is scalable to large-scale datasets. SSDH is simple and can be realized by a slight enhancement of an existing deep architecture for classification; yet it is effective and outperforms other hashing approaches on several benchmarks and large datasets. Compared with state-of-the-art approaches, SSDH achieves higher retrieval accuracy, while the classification performance is not sacrificed.",
"Cross-media hashing, which conducts cross-media retrieval by embedding data from different modalities into a common low-dimensional Hamming space, has attracted intensive attention in recent years. The existing cross-media hashing approaches only aim at learning hash functions to preserve the intra-modality and inter-modality correlations, but do not directly capture the underlying semantic information of the multi-modal data. We propose a discriminative coupled dictionary hashing (DCDH) method in this paper. In DCDH, the coupled dictionary for each modality is learned with side information (e.g., categories). As a result, the coupled dictionaries not only preserve the intra-similarity and inter-correlation among multi-modal data, but also contain dictionary atoms that are semantically discriminative (i.e., the data from the same category is reconstructed by the similar dictionary atoms). To perform fast cross-media retrieval, we learn hash functions which map data from the dictionary space to a low-dimensional Hamming space. Besides, we conjecture that a balanced representation is crucial in cross-media retrieval. We introduce multi-view features on the relatively weak'' modalities into DCDH and extend it to multi-view DCDH (MV-DCDH) in order to enhance their representation capability. The experiments on two real-world data sets show that our DCDH and MV-DCDH outperform the state-of-the-art methods significantly on cross-media retrieval.",
"Deep neural networks trained using a softmax layer at the top and the cross-entropy loss are ubiquitous tools for image classification. Yet, this does not naturally enforce intra-class similarity nor inter-class margin of the learned deep representations. To simultaneously achieve these two goals, different solutions have been proposed in the literature, such as the pairwise or triplet losses. However, these carry the extra task of selecting pairs or triplets, and the extra computational burden of computing and learning for many combinations of them. In this paper, we propose a plug-and-play loss term for deep networks that explicitly reduces intra-class variance and enforces inter-class margin simultaneously, in a simple and elegant geometric manner. For each class, the deep features are collapsed into a learned linear subspace, or union of them, and inter-class subspaces are pushed to be as orthogonal as possible. Our proposed Orthogonal Low-rank Embedding (OLA‰) does not require carefully crafting pairs or triplets of samples for training, and works standalone as a classification loss, being the first reported deep metric learning framework of its kind. Because of the improved margin between features of different classes, the resulting deep networks generalize better, are more discriminative, and more robust. We demonstrate improved classification performance in general object recognition, plugging the proposed loss term into existing off-the-shelf architectures. In particular, we show the advantage of the proposed loss in the small data model scenario, and we significantly advance the state-of-the-art on the Stanford STL-10 benchmark.",
"In this paper, we focus on improving the proposal classification stage in the object detection task and present implicit negative sub-categorization and sink diversion to lift the performance by strengthening loss function in this stage. First, based on the observation that the “background” class is generally very diverse and thus challenging to be handled as a single indiscriminative class in existing state-of-the-art methods, we propose to divide the background category into multiple implicit sub-categories to explicitly differentiate diverse patterns within it. Second, since the ground truth class inevitably has low-value probability scores for certain images, we propose to add a “sink” class and divert the probabilities of wrong classes to this class when necessary, such that the ground truth label will still have a higher probability than other wrong classes even though it has low probability output. Additionally, we propose to use dilated convolution, which is widely used in the semantic segmentation task, for efficient and valuable context information extraction. Extensive experiments on PASCAL VOC 2007 and 2012 data sets show that our proposed methods based on faster R-CNN implementation can achieve state-of-the-art mAPs, i.e., 84.1 , 82.6 , respectively, and obtain 2.5 improvement on ILSVRC DET compared with that of ResNet."
]
} |
1901.11259 | 2913092861 | Convolutional neural networks have been widely used in content-based image retrieval. To better deal with large-scale data, the deep hashing model is proposed as an effective method, which maps an image to a binary code that can be used for hashing search. However, most existing deep hashing models only utilize fine-level semantic labels or convert them to similar dissimilar labels for training. The natural semantic hierarchy structures are ignored in the training stage of the deep hashing model. In this paper, we present an effective algorithm to train a deep hashing model that can preserve a semantic hierarchy structure for large-scale image retrieval. Experiments on two datasets show that our method improves the fine-level retrieval performance. Meanwhile, our model achieves state-of-the-art results in terms of hierarchical retrieval. | Recently, the semantic hierarchy learning problem has been addressed in several works. The hierarchical semantic image retrieval model is proposed in @cite_14 . This work encodes hierarchy in semantic similarity. By combining the coarse and fine level labels, work in @cite_18 proves that the image classification performance can be improved. The similar idea has been shared in @cite_5 , which considers using a hierarchical training strategy to handle the face recognition task. Recent work in @cite_5 integrates the semantic relationship between class level into deep learning. The hierarchical similarity learning problem has also been addressed for deep hashing in SHDH @cite_21 . SHDH tackles the semantic hierarchy learning by proposing a weighted Hamming distance. However, SHDH also uses pair-wise relation, which has been proved not as efficient as using class-level labels @cite_17 @cite_12 . | {
"cite_N": [
"@cite_18",
"@cite_14",
"@cite_21",
"@cite_5",
"@cite_12",
"@cite_17"
],
"mid": [
"1581592866",
"1978962787",
"2963506586",
"2153419563",
"2170881581",
"2950940417"
],
"abstract": [
"In this work we propose a hierarchical approach for labeling semantic objects and regions in scenes. Our approach is reminiscent of early vision literature in that we use a decomposition of the image in order to encode relational and spatial information. In contrast to much existing work on structured prediction for scene understanding, we bypass a global probabilistic model and instead directly train a hierarchical inference procedure inspired by the message passing mechanics of some approximate inference procedures in graphical models. This approach mitigates both the theoretical and empirical difficulties of learning probabilistic models when exact inference is intractable. In particular, we draw from recent work in machine learning and break the complex inference process into a hierarchical series of simple machine learning subproblems. Each subproblem in the hierarchy is designed to capture the image and contextual statistics in the scene. This hierarchy spans coarse-to-fine regions and explicitly models the mixtures of semantic labels that may be present due to imperfect segmentation. To avoid cascading of errors and overfitting, we train the learning problems in sequence to ensure robustness to likely errors earlier in the inference sequence and leverage the stacking approach developed by",
"This paper addresses the problem of similar image retrieval, especially in the setting of large-scale datasets with millions to billions of images. The core novel contribution is an approach that can exploit prior knowledge of a semantic hierarchy. When semantic labels and a hierarchy relating them are available during training, significant improvements over the state of the art in similar image retrieval are attained. While some of this advantage comes from the ability to use additional information, experiments exploring a special case where no additional data is provided, show the new approach can still outperform OASIS [6], the current state of the art for similarity learning. Exploiting hierarchical relationships is most important for larger scale problems, where scalability becomes crucial. The proposed learning approach is fundamentally parallelizable and as a result scales more easily than previous work. An additional contribution is a novel hashing scheme (for bilinear similarity on vectors of probabilities, optionally taking into account hierarchy) that is able to reduce the computational cost of retrieval. Experiments are performed on Caltech256 and the larger ImageNet dataset.",
"Deep neural networks trained for classification have been found to learn powerful image representations, which are also often used for other tasks such as comparing images w.r.t. their visual similarity. However, visual similarity does not imply semantic similarity. In order to learn semantically discriminative features, we propose to map images onto class embeddings whose pair-wise dot products correspond to a measure of semantic similarity between classes. Such an embedding does not only improve image retrieval results, but could also facilitate integrating semantics for other tasks, e.g., novelty detection or few-shot learning. We introduce a deterministic algorithm for computing the class centroids directly based on prior world-knowledge encoded in a hierarchy of classes such as WordNet. Experiments on CIFAR-100, NABirds, and ImageNet show that our learned semantic image embeddings improve the semantic consistency of image retrieval results by a large margin.",
"We introduce an approach to learn discriminative visual representations while exploiting external semantic knowledge about object category relationships. Given a hierarchical taxonomy that captures semantic similarity between the objects, we learn a corresponding tree of metrics (ToM). In this tree, we have one metric for each non-leaf node of the object hierarchy, and each metric is responsible for discriminating among its immediate subcategory children. Specifically, a Mahalanobis metric learned for a given node must satisfy the appropriate (dis)similarity constraints generated only among its subtree members' training instances. To further exploit the semantics, we introduce a novel regularizer coupling the metrics that prefers a sparse disjoint set of features to be selected for each metric relative to its ancestor (supercategory) nodes' metrics. Intuitively, this reflects that visual cues most useful to distinguish the generic classes (e.g., feline vs. canine) should be different than those cues most useful to distinguish their component fine-grained classes (e.g., Persian cat vs. Siamese cat). We validate our approach with multiple image datasets using the WordNet taxonomy, show its advantages over alternative metric learning approaches, and analyze the meaning of attribute features selected by our algorithm.",
"We address the problem of cross-domain image retrieval, considering the following practical application: given a user photo depicting a clothing image, our goal is to retrieve the same or attribute-similar clothing items from online shopping stores. This is a challenging problem due to the large discrepancy between online shopping images, usually taken in ideal lighting pose background conditions, and user photos captured in uncontrolled conditions. To address this problem, we propose a Dual Attribute-aware Ranking Network (DARN) for retrieval feature learning. More specifically, DARN consists of two sub-networks, one for each domain, whose retrieval feature representations are driven by semantic attribute learning. We show that this attribute-guided learning is a key factor for retrieval accuracy improvement. In addition, to further align with the nature of the retrieval problem, we impose a triplet visual similarity constraint for learning to rank across the two subnetworks. Another contribution of our work is a large-scale dataset which makes the network learning feasible. We exploit customer review websites to crawl a large set of online shopping images and corresponding offline user photos with fine-grained clothing attributes, i.e., around 450,000 online shopping images and about 90,000 exact offline counterpart images of those online ones. All these images are collected from real-world consumer websites reflecting the diversity of the data modality, which makes this dataset unique and rare in the academic community. We extensively evaluate the retrieval performance of networks in different configurations. The top-20 retrieval accuracy is doubled when using the proposed DARN other than the current popular solution using pre-trained CNN features only (0.570 vs. 0.268).",
"We address the problem of cross-domain image retrieval, considering the following practical application: given a user photo depicting a clothing image, our goal is to retrieve the same or attribute-similar clothing items from online shopping stores. This is a challenging problem due to the large discrepancy between online shopping images, usually taken in ideal lighting pose background conditions, and user photos captured in uncontrolled conditions. To address this problem, we propose a Dual Attribute-aware Ranking Network (DARN) for retrieval feature learning. More specifically, DARN consists of two sub-networks, one for each domain, whose retrieval feature representations are driven by semantic attribute learning. We show that this attribute-guided learning is a key factor for retrieval accuracy improvement. In addition, to further align with the nature of the retrieval problem, we impose a triplet visual similarity constraint for learning to rank across the two sub-networks. Another contribution of our work is a large-scale dataset which makes the network learning feasible. We exploit customer review websites to crawl a large set of online shopping images and corresponding offline user photos with fine-grained clothing attributes, i.e., around 450,000 online shopping images and about 90,000 exact offline counterpart images of those online ones. All these images are collected from real-world consumer websites reflecting the diversity of the data modality, which makes this dataset unique and rare in the academic community. We extensively evaluate the retrieval performance of networks in different configurations. The top-20 retrieval accuracy is doubled when using the proposed DARN other than the current popular solution using pre-trained CNN features only (0.570 vs. 0.268)."
]
} |
1901.11448 | 2952848538 | The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings. | Multi-Domain Learning (MDL) MDL addresses training a single model capable of solving multiple datasets (domains). If the data is relatively small and the domains are similar, this sharing can lead to improved performance compared to training a separate model per domain @cite_25 . On the other hand, for diverse domains with large data MDL may under-perform a single model per domain; but is nonetheless is of interest due to the simplicity of a single model and its better memory scalability compared to a separate model per domain @cite_27 @cite_29 . We mention MDL here, because DG methods typically train on multiple source domains as per MDL -- but furthermore aim to generalise to novel held out domains. | {
"cite_N": [
"@cite_27",
"@cite_29",
"@cite_25"
],
"mid": [
"2786081970",
"2897482938",
"1703030490"
],
"abstract": [
"While domain adaptation has been actively researched in recent years, most theoretical results and algorithms focus on the single-source-single-target adaptation setting. Naive application of such algorithms on multiple source domain adaptation problem may lead to suboptimal solutions. We propose a new generalization bound for domain adaptation when there are multiple source domains with labeled instances and one target domain with unlabeled instances. Compared with existing bounds, the new bound does not require expert knowledge about the target distribution, nor the optimal combination rule for multisource domains. Interestingly, our theory also leads to an efficient learning strategy using adversarial neural networks: we show how to interpret it as learning feature representations that are invariant to the multiple domain shifts while still being discriminative for the learning task. To this end, we propose two models, both of which we call multisource domain adversarial networks (MDANs): the first model optimizes directly our bound, while the second model is a smoothed approximation of the first one, leading to a more data-efficient and task-adaptive model. The optimization tasks of both models are minimax saddle point problems that can be optimized by adversarial training. To demonstrate the effectiveness of MDANs, we conduct extensive experiments showing superior adaptation performance on three real-world datasets: sentiment analysis, digit classification, and vehicle counting.",
"Domain adaptation aims to train a model on labeled data from a source domain while minimizing test error on a target domain. Most of existing domain adaptation methods only focus on reducing domain shift of single-modal data. In this paper, we consider a new problem of multimodal domain adaptation and propose a unified framework to solve it. The proposed multimodal domain adaptation neural networks(MDANN) consist of three important modules. (1) A covariant multimodal attention is designed to learn a common feature representation for multiple modalities. (2) A fusion module adaptively fuses attended features of different modalities. (3) Hybrid domain constraints are proposed to comprehensively learn domain-invariant features by constraining single modal features, fused features, and attention scores. Through jointly attending and fusing under an adversarial objective, the most discriminative and domain-adaptive parts of the features are adaptively fused together. Extensive experimental results on two real-world cross-domain applications (emotion recognition and cross-media retrieval) demonstrate the effectiveness of the proposed method.",
"In this paper, we provide a new neural-network based perspective on multi-task learning (MTL) and multi-domain learning (MDL). By introducing the concept of a semantic descriptor, this framework unifies MDL and MTL as well as encompassing various classic and recent MTL MDL algorithms by interpreting them as different ways of constructing semantic descriptors. Our interpretation provides an alternative pipeline for zero-shot learning (ZSL), where a model for a novel class can be constructed without training data. Moreover, it leads to a new and practically relevant problem setting of zero-shot domain adaptation (ZSDA), which is the analogous to ZSL but for novel domains: A model for an unseen domain can be generated by its semantic descriptor. Experiments across this range of problems demonstrate that our framework outperforms a variety of alternatives."
]
} |
1901.11448 | 2952848538 | The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings. | Most existing DG approaches can be split into three categories: feature-based methods, classifier-based methods, and data augmentation methods. Feature-based methods: These aim to generate a domain-invariant representation. For example where the distance between the empirical distributions of the source and target examples is minimized @cite_2 @cite_22 @cite_3 . Classifier-based methods: Some aim to enhance generalisation by fusing multiple sub-classifiers learned from source domains @cite_21 @cite_20 @cite_34 , and others learn an improved classifier regularizer using source samples -- notably the recently proposed MetaReg @cite_8 . Data augmentation methods: CrossGrad @cite_19 generates provides domain-guided perturbations of input instances, which are then used to train a more robust model. defines an adaptive data augmentation scheme by appending adversarial examples at each iteration. Our approach falls into the feature-based category, but meta-learns a feature-critic network to train a robust shared feature extractor. | {
"cite_N": [
"@cite_22",
"@cite_8",
"@cite_21",
"@cite_3",
"@cite_19",
"@cite_2",
"@cite_34",
"@cite_20"
],
"mid": [
"2767382337",
"2964139811",
"2962687275",
"2787223504",
"2798681837",
"2771773135",
"2770856226",
"2964057616"
],
"abstract": [
"We present a method for transferring neural representations from label-rich source domains to unlabeled target domains. Recent adversarial methods proposed for this task learn to align features across domains by fooling a special domain critic network. However, a drawback of this approach is that the critic simply labels the generated features as in-domain or not, without considering the boundaries between classes. This can lead to ambiguous features being generated near class boundaries, reducing target classification accuracy. We propose a novel approach, Adversarial Dropout Regularization (ADR), to encourage the generator to output more discriminative features for the target domain. Our key idea is to replace the critic with one that detects non-discriminative features, using dropout on the classifier network. The generator then learns to avoid these areas of the feature space and thus creates better features. We apply our ADR approach to the problem of unsupervised domain adaptation for image classification and semantic segmentation tasks, and demonstrate significant improvement over the state of the art. We also show that our approach can be used to train Generative Adversarial Networks for semi-supervised learning.",
"We present a method for transferring neural representations from label-rich source domains to unlabeled target domains. Recent adversarial methods proposed for this task learn to align features across domains by fooling a special domain critic network. However, a drawback of this approach is that the critic simply labels the generated features as in-domain or not, without considering the boundaries between classes. This can lead to ambiguous features being generated near class boundaries, reducing target classification accuracy. We propose a novel approach, Adversarial Dropout Regularization (ADR), to encourage the generator to output more discriminative features for the target domain. Our key idea is to replace the critic with one that detects non-discriminative features, using dropout on the classifier network. The generator then learns to avoid these areas of the feature space and thus creates better features. We apply our ADR approach to the problem of unsupervised domain adaptation for image classification and semantic segmentation tasks, and demonstrate significant improvement over the state of the art. We also show that our approach can be used to train Generative Adversarial Networks for semi-supervised learning.",
"In this work, we present a method for unsupervised domain adaptation. Many adversarial learning methods train domain classifier networks to distinguish the features as either a source or target and train a feature generator network to mimic the discriminator. Two problems exist with these methods. First, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes. Therefore, a trained generator can generate ambiguous features near class boundaries. Second, these methods aim to completely match the feature distributions between different domains, which is difficult because of each domain's characteristics. To solve these problems, we introduce a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to maximize the discrepancy between two classifiers' outputs to detect target samples that are far from the support of the source. A feature generator learns to generate target features near the support to minimize the discrepancy. Our method outperforms other methods on several datasets of image classification and semantic segmentation. The codes are available at https: github.com mil-tokyo MCD_DA",
"We propose in this paper a new approach to train the Generative Adversarial Nets (GANs) with a mixture of generators to overcome the mode collapsing problem. The main intuition is to employ multiple generators, instead of using a single one as in the original GAN. The idea is simple, yet proven to be extremely effective at covering diverse data modes, easily overcoming the mode collapsing problem and delivering state-of-the-art results. A minimax formulation was able to establish among a classifier, a discriminator, and a set of generators in a similar spirit with GAN. Generators create samples that are intended to come from the same distribution as the training data, whilst the discriminator determines whether samples are true data or generated by generators, and the classifier specifies which generator a sample comes from. The distinguishing feature is that internal samples are created from multiple generators, and then one of them will be randomly selected as final output similar to the mechanism of a probabilistic mixture model. We term our method Mixture Generative Adversarial Nets (MGAN). We develop theoretical analysis to prove that, at the equilibrium, the Jensen-Shannon divergence (JSD) between the mixture of generators’ distributions and the empirical data distribution is minimal, whilst the JSD among generators’ distributions is maximal, hence effectively avoiding the mode collapsing problem. By utilizing parameter sharing, our proposed model adds minimal computational cost to the standard GAN, and thus can also efficiently scale to large-scale datasets. We conduct extensive experiments on synthetic 2D data and natural image databases (CIFAR-10, STL-10 and ImageNet) to demonstrate the superior performance of our MGAN in achieving state-of-the-art Inception scores over latest baselines, generating diverse and appealing recognizable objects at different resolutions, and specializing in capturing different types of objects by the generators.",
"In this paper, we propose a new unsupervised domain adaptation approach called Collaborative and Adversarial Network (CAN) through domain-collaborative and domain-adversarial training of neural networks. We add several domain classifiers on multiple CNN feature extraction blocks1, in which each domain classifier is connected to the hidden representations from one block and one loss function is defined based on the hidden presentation and the domain labels (e.g., source and target). We design a new loss function by integrating the losses from all blocks in order to learn domain informative representations from lower blocks through collaborative learning and learn domain uninformative representations from higher blocks through adversarial learning. We further extend our CAN method as Incremental CAN (iCAN), in which we iteratively select a set of pseudo-labelled target samples based on the image classifier and the last domain classifier from the previous training epoch and re-train our CAN model by using the enlarged training set. Comprehensive experiments on two benchmark datasets Office and ImageCLEF-DA clearly demonstrate the effectiveness of our newly proposed approaches CAN and iCAN for unsupervised domain adaptation.",
"In this work, we present a method for unsupervised domain adaptation (UDA), where we aim to transfer knowledge from a label-rich domain (i.e., a source domain) to an unlabeled domain (i.e., a target domain). Many adversarial learning methods have been proposed for this task. These methods train domain classifier networks (i.e., a discriminator) to discriminate distinguish the features as either a source or target and train a feature generator network to mimic the discriminator.However, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes. Therefore, a trained generator can generate ambiguous features near class boundaries. To solve the problem, we propose a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to utilize task-specific classifiers as discriminators that try to detect target samples that are far from the support of the source. A feature generator learns to generate target features inside the support to fool the classifiers. Since the generator uses feedback from task-specific classifiers, it avoids generating target features near class boundaries. Our method outperforms other methods on several datasets of image classification and semantic segmentation.",
"Recent works showed that Generative Adversarial Networks (GANs) can be successfully applied in unsupervised domain adaptation, where, given a labeled source dataset and an unlabeled target dataset, the goal is to train powerful classifiers for the target samples. In particular, it was shown that a GAN objective function can be used to learn target features indistinguishable from the source ones. In this work, we extend this framework by (i) forcing the learned feature extractor to be domain-invariant, and (ii) training it through data augmentation in the feature space, namely performing feature augmentation. While data augmentation in the image space is a well established technique in deep learning, feature augmentation has not yet received the same level of attention. We accomplish it by means of a feature generator trained by playing the GAN minimax game against source features. Results show that both enforcing domain-invariance and performing feature augmentation lead to superior or comparable performance to state-of-the-art results in several unsupervised domain adaptation benchmarks.",
"Recent works showed that Generative Adversarial Networks (GANs) can be successfully applied in unsupervised domain adaptation, where, given a labeled source dataset and an unlabeled target dataset, the goal is to train powerful classifiers for the target samples. In particular, it was shown that a GAN objective function can be used to learn target features indistinguishable from the source ones. In this work, we extend this framework by (i) forcing the learned feature extractor to be domain-invariant, and (ii) training it through data augmentation in the feature space, namely performing feature augmentation. While data augmentation in the image space is a well established technique in deep learning, feature augmentation has not yet received the same level of attention. We accomplish it by means of a feature generator trained by playing the GAN minimax game against source features. Results show that both enforcing domain-invariance and performing feature augmentation lead to superior or comparable performance to state-of-the-art results in several unsupervised domain adaptation benchmarks."
]
} |
1901.11448 | 2952848538 | The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings. | Few studies have considered the heterogeneous DG setting, where the domains do not share the same label space. In this setting, we do not expect the to generalize directly to the target domain (impossible due to the change in label space), but we do aim to improve the robustness of a source-domain trained in terms of its generalisation to successfully represent a novel problem. Most existing DG methods cannot be applied here besides Domain Adaptive Neural Networks @cite_5 and CrossGrad @cite_19 . We show how to modify MetaReg @cite_8 and Reptile @cite_37 algorithms to address this DG setting. The most relevant benchmark is the Visual Decathlon (VD) @cite_27 . The VD benchmark was proposed to evaluate multi-domain and lifelong @cite_31 learning. VD competitors should originally learn a model covering all ten domains, with low parameter growth. We re-purpose the VD benchmark for DG evaluation. In this case a model trained on the six largest datasets in VD should produce a feature which provides a general and robust enough encoding to allow the four smaller data-sets to be classified with a simple shallow classifier. | {
"cite_N": [
"@cite_37",
"@cite_8",
"@cite_19",
"@cite_27",
"@cite_5",
"@cite_31"
],
"mid": [
"2763549966",
"2951454049",
"2363300041",
"2557626841",
"2767382337",
"2964139811"
],
"abstract": [
"The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied to an unseen domain. Domain generalization (DG) has a clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition in sketch images, which are distinctly more abstract and rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing on alleviating the dataset bias where both problems of domain distinctiveness and data sparsity can be minimal. We argue that these benchmarks are overly straightforward, and show that simple deep learning baselines perform surprisingly well on them. In this paper, we make two main contributions: Firstly, we build upon the favorable domain shift-robust properties of deep learning methods, and develop a low-rank parameterized CNN model for end-to-end DG learning. Secondly, we develop a DG benchmark dataset covering photo, sketch, cartoon and painting domains. This is both more practically relevant, and harder (bigger domain shift) than existing benchmarks. The results show that our method outperforms existing DG alternatives, and our dataset provides a more significant DG challenge to drive future research.",
"The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied to an unseen domain. Domain generalization (DG) has a clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition in sketch images, which are distinctly more abstract and rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing on alleviating the dataset bias where both problems of domain distinctiveness and data sparsity can be minimal. We argue that these benchmarks are overly straightforward, and show that simple deep learning baselines perform surprisingly well on them. In this paper, we make two main contributions: Firstly, we build upon the favorable domain shift-robust properties of deep learning methods, and develop a low-rank parameterized CNN model for end-to-end DG learning. Secondly, we develop a DG benchmark dataset covering photo, sketch, cartoon and painting domains. This is both more practically relevant, and harder (bigger domain shift) than existing benchmarks. The results show that our method outperforms existing DG alternatives, and our dataset provides a more significant DG challenge to drive future research.",
"One of the serious challenges in computer vision and image classification is learning an accurate classifier for a new unlabeled image dataset, considering that there is no available labeled training data. Transfer learning and domain adaptation are two outstanding solutions that tackle this challenge by employing available datasets, even with significant difference in distribution and properties, and transfer the knowledge from a related domain to the target domain. The main difference between these two solutions is their primary assumption about change in marginal and conditional distributions where transfer learning emphasizes on problems with same marginal distribution and different conditional distribution, and domain adaptation deals with opposite conditions. Most prior works have exploited these two learning strategies separately for domain shift problem where training and test sets are drawn from different distributions. In this paper, we exploit joint transfer learning and domain adaptation to cope with domain shift problem in which the distribution difference is significantly large, particularly vision datasets. We therefore put forward a novel transfer learning and domain adaptation approach, referred to as visual domain adaptation (VDA). Specifically, VDA reduces the joint marginal and conditional distributions across domains in an unsupervised manner where no label is available in test set. Moreover, VDA constructs condensed domain invariant clusters in the embedding representation to separate various classes alongside the domain transfer. In this work, we employ pseudo target labels refinement to iteratively converge to final solution. Employing an iterative procedure along with a novel optimization problem creates a robust and effective representation for adaptation across domains. Extensive experiments on 16 real vision datasets with different difficulties verify that VDA can significantly outperform state-of-the-art methods in image classification problem.",
"In this paper, we propose an approach to the domain adaptation, dubbed Second-or Higher-order Transfer of Knowledge (So-HoT), based on the mixture of alignments of second-or higher-order scatter statistics between the source and target domains. The human ability to learn from few labeled samples is a recurring motivation in the literature for domain adaptation. Towards this end, we investigate the supervised target scenario for which few labeled target training samples per category exist. Specifically, we utilize two CNN streams: the source and target networks fused at the classifier level. Features from the fully connected layers fc7 of each network are used to compute second-or even higher-order scatter tensors, one per network stream per class. As the source and target distributions are somewhat different despite being related, we align the scatters of the two network streams of the same class (within-class scatters) to a desired degree with our bespoke loss while maintaining good separation of the between-class scatters. We train the entire network in end-to-end fashion. We provide evaluations on the standard Office benchmark (visual domains) and RGB-D combined with Caltech256 (depth-to-rgb transfer). We attain state-of-the-art results.",
"We present a method for transferring neural representations from label-rich source domains to unlabeled target domains. Recent adversarial methods proposed for this task learn to align features across domains by fooling a special domain critic network. However, a drawback of this approach is that the critic simply labels the generated features as in-domain or not, without considering the boundaries between classes. This can lead to ambiguous features being generated near class boundaries, reducing target classification accuracy. We propose a novel approach, Adversarial Dropout Regularization (ADR), to encourage the generator to output more discriminative features for the target domain. Our key idea is to replace the critic with one that detects non-discriminative features, using dropout on the classifier network. The generator then learns to avoid these areas of the feature space and thus creates better features. We apply our ADR approach to the problem of unsupervised domain adaptation for image classification and semantic segmentation tasks, and demonstrate significant improvement over the state of the art. We also show that our approach can be used to train Generative Adversarial Networks for semi-supervised learning.",
"We present a method for transferring neural representations from label-rich source domains to unlabeled target domains. Recent adversarial methods proposed for this task learn to align features across domains by fooling a special domain critic network. However, a drawback of this approach is that the critic simply labels the generated features as in-domain or not, without considering the boundaries between classes. This can lead to ambiguous features being generated near class boundaries, reducing target classification accuracy. We propose a novel approach, Adversarial Dropout Regularization (ADR), to encourage the generator to output more discriminative features for the target domain. Our key idea is to replace the critic with one that detects non-discriminative features, using dropout on the classifier network. The generator then learns to avoid these areas of the feature space and thus creates better features. We apply our ADR approach to the problem of unsupervised domain adaptation for image classification and semantic segmentation tasks, and demonstrate significant improvement over the state of the art. We also show that our approach can be used to train Generative Adversarial Networks for semi-supervised learning."
]
} |
1901.11448 | 2952848538 | The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings. | Meta-Learning Meta-learning (a.k.a. learning to learn, @cite_15 @cite_13 ) has received resurgence in interest recently with applications in few-shot learning @cite_9 @cite_40 and beyond @cite_16 . In few-shot meta-learning, a common strategy is to simulate the few-shot learning scenario by randomly drawing few-shot train test episodes from the full training set. Training the network to solve such episodes tunes it to perform well at few-shot learning. We adapt this episodic training strategy by creating virtual training and testing splits of our source domains in each mini-batch. | {
"cite_N": [
"@cite_9",
"@cite_40",
"@cite_15",
"@cite_16",
"@cite_13"
],
"mid": [
"2903334135",
"2807352135",
"2964078140",
"2786928087",
"2742093937"
],
"abstract": [
"Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. The key idea is to leverage a large number of similar few-shot tasks in order to learn how to adapt a base-learner to a new task for which only a few labeled samples are available. As deep neural networks (DNNs) tend to overfit using a few samples only, meta-learning typically uses shallow neural networks (SNNs), thus limiting its effectiveness. In this paper we propose a novel few-shot learning method called meta-transfer learning (MTL) which learns to adapt a deep NN for few shot learning tasks. Specifically, \"meta\" refers to training multiple tasks, and \"transfer\" is achieved by learning scaling and shifting functions of DNN weights for each task. In addition, we introduce the hard task (HT) meta-batch scheme as an effective learning curriculum for MTL. We conduct experiments using (5-class, 1-shot) and (5-class, 5-shot) recognition tasks on two challenging few-shot learning benchmarks: miniImageNet and Fewshot-CIFAR100. Extensive comparisons to related works validate that our meta-transfer learning approach trained with the proposed HT meta-batch scheme achieves top performance. An ablation study also shows that both components contribute to fast convergence and high accuracy.",
"Meta-learning for few-shot learning entails acquiring a prior over previous tasks and experiences, such that new tasks be learned from small amounts of data. However, a critical challenge in few-shot learning is task ambiguity: even when a powerful prior can be meta-learned from a large number of prior tasks, a small dataset for a new task can simply be too ambiguous to acquire a single model (e.g., a classifier) for that task that is accurate. In this paper, we propose a probabilistic meta-learning algorithm that can sample models for a new task from a model distribution. Our approach extends model-agnostic meta-learning, which adapts to new tasks via gradient descent, to incorporate a parameter distribution that is trained via a variational lower bound. At meta-test time, our algorithm adapts via a simple procedure that injects noise into gradient descent, and at meta-training time, the model is trained such that this stochastic adaptation procedure produces samples from the approximate model posterior. Our experimental results show that our method can sample plausible classifiers and regressors in ambiguous few-shot learning problems.",
"Meta-learning for few-shot learning entails acquiring a prior over previous tasks and experiences, such that new tasks be learned from small amounts of data. However, a critical challenge in few-shot learning is task ambiguity: even when a powerful prior can be meta-learned from a large number of prior tasks, a small dataset for a new task can simply be too ambiguous to acquire a single model (e.g., a classifier) for that task that is accurate. In this paper, we propose a probabilistic meta-learning algorithm that can sample models for a new task from a model distribution. Our approach extends model-agnostic meta-learning, which adapts to new tasks via gradient descent, to incorporate a parameter distribution that is trained via a variational lower bound. At meta-test time, our algorithm adapts via a simple procedure that injects noise into gradient descent, and at meta-training time, the model is trained such that this stochastic adaptation procedure produces samples from the approximate model posterior. Our experimental results show that our method can sample plausible classifiers and regressors in ambiguous few-shot learning problems.",
"Few-shot learning remains challenging for meta-learning that learns a learning algorithm (meta-learner) from many related tasks. In this work, we argue that this is due to the lack of a good representation for meta-learning, and propose deep meta-learning to integrate the representation power of deep learning into meta-learning. The framework is composed of three modules, a concept generator, a meta-learner, and a concept discriminator, which are learned jointly. The concept generator, e.g. a deep residual net, extracts a representation for each instance that captures its high-level concept, on which the meta-learner performs few-shot learning, and the concept discriminator recognizes the concepts. By learning to learn in the concept space rather than in the complicated instance space, deep meta-learning can substantially improve vanilla meta-learning, which is demonstrated on various few-shot image recognition problems. For example, on 5-way-1-shot image recognition on CIFAR-100 and CUB-200, it improves Matching Nets from 50.53 and 56.53 to 58.18 and 63.47 , improves MAML from 49.28 and 50.45 to 56.65 and 64.63 , and improves Meta-SGD from 53.83 and 53.34 to 61.62 and 66.95 , respectively.",
"Few-shot learning is challenging for learning algorithms that learn each task in isolation and from scratch. In contrast, meta-learning learns from many related tasks a meta-learner that can learn a new task more accurately and faster with fewer examples, where the choice of meta-learners is crucial. In this paper, we develop Meta-SGD, an SGD-like, easily trainable meta-learner that can initialize and adapt any differentiable learner in just one step, on both supervised learning and reinforcement learning. Compared to the popular meta-learner LSTM, Meta-SGD is conceptually simpler, easier to implement, and can be learned more efficiently. Compared to the latest meta-learner MAML, Meta-SGD has a much higher capacity by learning to learn not just the learner initialization, but also the learner update direction and learning rate, all in a single meta-learning process. Meta-SGD shows highly competitive performance for few-shot learning on regression, classification, and reinforcement learning."
]
} |
1901.11448 | 2952848538 | The well known domain shift issue causes model performance to degrade when deployed to a new target domain with different statistics to training. Domain adaptation techniques alleviate this, but need some instances from the target domain to drive adaptation. Domain generalisation is the recently topical problem of learning a model that generalises to unseen domains out of the box, and various approaches aim to train a domain-invariant feature extractor, typically by adding some manually designed losses. In this work, we propose a learning to learn approach, where the auxiliary loss that helps generalisation is itself learned. Beyond conventional domain generalisation, we consider a more challenging setting of heterogeneous domain generalisation, where the unseen domains do not share label space with the seen ones, and the goal is to train a feature representation that is useful off-the-shelf for novel data and novel categories. Experimental evaluation demonstrates that our method outperforms state-of-the-art solutions in both settings. | A few methods have applied related episodic meta-learning strategies in DG @cite_2 @cite_8 . MLDG @cite_2 defined a heuristic gradient descent update rule based on the gradients of the simulated training and testing domains. MetaReg @cite_8 trains the weights of the 's regulariser so as to produce a more general classifier for a fixed feature extractor. In contrast, our produces a more general that can be used with any classifier. This is achieved by simultaneously learning an auxiliary loss function @cite_24 (i.e., the critic network) that helps to train the feature extractor for improved domain invariance. | {
"cite_N": [
"@cite_24",
"@cite_8",
"@cite_2"
],
"mid": [
"2737691244",
"2787387965",
"2964227899"
],
"abstract": [
"In this paper, we study the problem of training large-scale face identification model with imbalanced training data. This problem naturally exists in many real scenarios including large-scale celebrity recognition, movie actor annotation, etc. Our solution contains two components. First, we build a face feature extraction model, and improve its performance, especially for the persons with very limited training samples, by introducing a regularizer to the cross entropy loss for the multi-nomial logistic regression (MLR) learning. This regularizer encourages the directions of the face features from the same class to be close to the direction of their corresponding classification weight vector in the logistic regression. Second, we build a multi-class classifier using MLR on top of the learned face feature extraction model. Since the standard MLR has poor generalization capability for the one-shot classes even if these classes have been oversampled, we propose a novel supervision signal called underrepresented-classes promotion loss, which aligns the norms of the weight vectors of the one-shot classes (a.k.a. underrepresented-classes) to those of the normal classes. In addition to the original cross entropy loss, this new loss term effectively promotes the underrepresented classes in the learned model and leads to a remarkable improvement in face recognition performance. We test our solution on the MS-Celeb-1M low-shot learning benchmark task. Our solution recognizes 94.89 of the test images at the precision of 99 for the one-shot classes. To the best of our knowledge, this is the best performance among all the published methods using this benchmark task with the same setup, including all the participants in the recent MS-Celeb-1M challenge at ICCV 2017.",
"We propose a met alearning approach for learning gradient-based reinforcement learning (RL) algorithms. The idea is to evolve a differentiable loss function, such that an agent, which optimizes its policy to minimize this loss, will achieve high rewards. The loss is parametrized via temporal convolutions over the agent's experience. Because this loss is highly flexible in its ability to take into account the agent's history, it enables fast task learning. Empirical results show that our evolved policy gradient algorithm (EPG) achieves faster learning on several randomized environments compared to an off-the-shelf policy gradient method. We also demonstrate that EPG's learned loss can generalize to out-of-distribution test time tasks, and exhibits qualitatively different behavior from other popular met alearning algorithms.",
"We propose a met alearning approach for learning gradient-based reinforcement learning (RL) algorithms. The idea is to evolve a differentiable loss function, such that an agent, which optimizes its policy to minimize this loss, will achieve high rewards. The loss is parametrized via temporal convolutions over the agent's experience. Because this loss is highly flexible in its ability to take into account the agent's history, it enables fast task learning. Empirical results show that our evolved policy gradient algorithm (EPG) achieves faster learning on several randomized environments compared to an off-the-shelf policy gradient method. We also demonstrate that EPG's learned loss can generalize to out-of-distribution test time tasks, and exhibits qualitatively different behavior from other popular met alearning algorithms."
]
} |
1901.11462 | 2952666417 | Conversational agents have begun to rise both in the academic (in terms of research) and commercial (in terms of applications) world. This paper investigates the task of building a non-goal driven conversational agent, using neural network generative models and analyzes how the conversation context is handled. It compares a simpler Encoder-Decoder with a Hierarchical Recurrent Encoder-Decoder architecture, which includes an additional module to model the context of the conversation using previous utterances information. We found that the hierarchical model was able to extract relevant context information and include them in the generation of the output. However, it performed worse (35-40 ) than the simple Encoder-Decoder model regarding both grammatically correct output and meaningful response. Despite these results, experiments demonstrate how conversations about similar topics appear close to each other in the context space due to the increased frequency of specific topic-related words, thus leaving promising directions for future research and how the context of a conversation can be exploited. | The traditional approach for Conversational Agents follows a modular approach, dividing the process into three modules: a Natural Language Understanding (NLU) unit, a Dialogue Manager and a Natural Language Generation module (NLG). The NLU module will process the input and extract useful information. This information is then used by the Dialogue Manager to update internal states, send a query to a knowledge-based system, or simply follow precoded instructions. Finally, the NLG will use the information from the Dialogue Manager to generate the output sentence. The simplest technique used for NLU is to spot certain keywords in the input, often working together with a script-based Dialogue Manager. However, throughout the years there have been many attempts to improve the NLU unit to better extract text information, using techniques including statistical modeling of language @cite_2 , skip-gram models @cite_4 and, more recently, deep neural networks @cite_3 . Eventually, with the rise of Deep Learning in recent years, Dialogue Systems research has mainly focused on end-to-end models, capable of including all 3 modules in a single deep neural network, trained on a large dataset. One end-to-end RNN architecture proved particularly successful in recent years is the Encoder-Decoder. | {
"cite_N": [
"@cite_4",
"@cite_3",
"@cite_2"
],
"mid": [
"2593751037",
"2806935606",
"2410983263"
],
"abstract": [
"In this paper, we construct and train end-to-end neural network-based dialogue systems using an updated version of the recent Ubuntu Dialogue Corpus, a dataset containing almost 1 million multi-turn dialogues, with a total of over 7 million utterances and 100 million words. This dataset is interesting because of its size, long context lengths, and technical nature; thus, it can be used to train large models directly from data with minimal feature engineering, which can be both time consuming and expensive. We provide baselines in two different environments: one where models are trained to maximize the log-likelihood of a generated utterance conditioned on the context of the conversation, and one where models are trained to select the correct next response from a list of candidate responses. These are both evaluated on a recall task that we call Next Utterance Classification (NUC), as well as other generation-specific metrics. Finally, we provide a qualitative error analysis to help determine the most promising directions for future research on the Ubuntu Dialogue Corpus, and for end-to-end dialogue systems in general.",
"We propose an adversarial learning approach for generating multi-turn dialogue responses. Our proposed framework, hredGAN, is based on conditional generative adversarial networks (GANs). The GAN's generator is a modified hierarchical recurrent encoder-decoder network (HRED) and the discriminator is a word-level bidirectional RNN that shares context and word embeddings with the generator. During inference, noise samples conditioned on the dialogue history are used to perturb the generator's latent space to generate several possible responses. The final response is the one ranked best by the discriminator. The hredGAN shows improved performance over existing methods: (1) it generalizes better than networks trained using only the log-likelihood criterion, and (2) it generates longer, more informative and more diverse responses with high utterance and topic relevance even with limited training data. This improvement is demonstrated on the Movie triples and Ubuntu dialogue datasets using both automatic and human evaluations.",
"Recent neural models of dialogue generation offer great promise for generating responses for conversational agents, but tend to be shortsighted, predicting utterances one at a time while ignoring their influence on future outcomes. Modeling the future direction of a dialogue is crucial to generating coherent, interesting dialogues, a need which led traditional NLP models of dialogue to draw on reinforcement learning. In this paper, we show how to integrate these goals, applying deep reinforcement learning to model future reward in chatbot dialogue. The model simulates dialogues between two virtual agents, using policy gradient methods to reward sequences that display three useful conversational properties: informativity (non-repetitive turns), coherence, and ease of answering (related to forward-looking function). We evaluate our model on diversity, length as well as with human judges, showing that the proposed algorithm generates more interactive responses and manages to foster a more sustained conversation in dialogue simulation. This work marks a first step towards learning a neural conversational model based on the long-term success of dialogues."
]
} |
1901.11462 | 2952666417 | Conversational agents have begun to rise both in the academic (in terms of research) and commercial (in terms of applications) world. This paper investigates the task of building a non-goal driven conversational agent, using neural network generative models and analyzes how the conversation context is handled. It compares a simpler Encoder-Decoder with a Hierarchical Recurrent Encoder-Decoder architecture, which includes an additional module to model the context of the conversation using previous utterances information. We found that the hierarchical model was able to extract relevant context information and include them in the generation of the output. However, it performed worse (35-40 ) than the simple Encoder-Decoder model regarding both grammatically correct output and meaningful response. Despite these results, experiments demonstrate how conversations about similar topics appear close to each other in the context space due to the increased frequency of specific topic-related words, thus leaving promising directions for future research and how the context of a conversation can be exploited. | The use of Encoder-Decoder architectures for natural language processing was first proposed as a solution for text translation, in 2014 by @cite_7 . From then on, the architecture has been applied to many other tasks, including conversational agents @cite_6 . However, generating responses was found to be considerably more difficult than translating between languages, probably due to a broader range of possible correct answers to any given input. A limitation of Encoder-Decoder models to produce meaningful conversations is the fact that any output is only influenced by the latest question. Thus, important factors are ignored, such as the context of the conversation, the speaker, and information provided in previous inputs. In 2015, proposed an updated version of the Encoder-Decoder architecture, called Hierarchical Recurrent Encoder Decoder @cite_8 (HRED), originally used for query suggestions. In their paper, they demonstrate that the architecture is capable of using context information extracted by previous queries to generate more appropriate query suggestions. This paper will attempt to apply such architecture to a dialogue system. | {
"cite_N": [
"@cite_6",
"@cite_7",
"@cite_8"
],
"mid": [
"1993378086",
"2280798142",
"2752047430"
],
"abstract": [
"Users may strive to formulate an adequate textual query for their information need. Search engines assist the users by presenting query suggestions. To preserve the original search intent, suggestions should be context-aware and account for the previous queries issued by the user. Achieving context awareness is challenging due to data sparsity. We present a novel hierarchical recurrent encoder-decoder architecture that makes possible to account for sequences of previous queries of arbitrary lengths. As a result, our suggestions are sensitive to the order of queries in the context while avoiding data sparsity. Additionally, our model can suggest for rare, or long-tail, queries. The produced suggestions are synthetic and are sampled one word at a time, using computationally cheap decoding techniques. This is in contrast to current synthetic suggestion models relying upon machine learning pipelines and hand-engineered feature sets. Results show that our model outperforms existing context-aware approaches in a next query prediction setting. In addition to query suggestion, our architecture is general enough to be used in a variety of other applications.",
"In this work, we cast text summarization as a sequence-to-sequence problem and apply the attentional encoder-decoder RNN that has been shown to be successful for Machine Translation ( (2014)). Our experiments show that the proposed architecture significantly outperforms the state-of-the art model of (2015) on the Gigaword dataset without any additional tuning. We also propose additional extensions to the standard architecture, which we show contribute to further improvement in performance.",
"The encoder-decoder framework has achieved promising progress for many sequence generation tasks, including machine translation, text summarization, dialog system, image captioning, etc. Such a framework adopts an one-pass forward process while decoding and generating a sequence, but lacks the deliberation process: A generated sequence is directly used as final output without further polishing. However, deliberation is a common behavior in human's daily life like reading news and writing papers articles books. In this work, we introduce the deliberation process into the encoder-decoder framework and propose deliberation networks for sequence generation. A deliberation network has two levels of decoders, where the first-pass decoder generates a raw sequence and the second-pass decoder polishes and refines the raw sentence with deliberation. Since the second-pass deliberation decoder has global information about what the sequence to be generated might be, it has the potential to generate a better sequence by looking into future words in the raw sentence. Experiments on neural machine translation and text summarization demonstrate the effectiveness of the proposed deliberation networks. On the WMT 2014 English-to-French translation task, our model establishes a new state-of-the-art BLEU score of 41.5."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | As might be expected, filter-based FS algorithms have asymptotic complexities that depend on the number of features and or instances in a dataset. Many algorithms, such as the CFS, have quadratic complexities, while the most frequently used algorithms have at least linear complexities @cite_26 . This is why, in recent years, many attempts have been made to achieve more scalable FS methods. In what follows, we analyse recent work on the design of new scalable FS methods according to parallelization approaches: (i) search-oriented, (ii) dataset-split-oriented, or (iii) filter-oriented. | {
"cite_N": [
"@cite_26"
],
"mid": [
"2952294252"
],
"abstract": [
"This paper describes a novel approach of packing sparse convolutional neural networks for their efficient systolic array implementations. By combining subsets of columns in the original filter matrix associated with a convolutional layer, we increase the utilization efficiency of the systolic array substantially (e.g., 4x) due to the increased density of nonzeros in the resulting packed filter matrix. In combining columns, for each row, all filter weights but one with the largest magnitude are pruned. We retrain the remaining weights to preserve high accuracy. We demonstrate that in mitigating data privacy concerns the retraining can be accomplished with only fractions of the original dataset (e.g., 10 for CIFAR-10). We study the effectiveness of this joint optimization for both high utilization and classification accuracy with ASIC and FPGA designs based on efficient bit-serial implementations of multiplier-accumulators. We present analysis and empirical evidence on the superior performance of our column combining approach against prior arts under metrics such as energy efficiency (3x) and inference latency (12x)."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | @cite_8 developed parallel versions of three forward-search-based FS algorithms, where a wrapper with a logistic regression classifier is used to guide a search parallelized using the MapReduce model. | {
"cite_N": [
"@cite_8"
],
"mid": [
"2573038016"
],
"abstract": [
"Abstract In this paper we present PFDCMSS (Parallel Forward Decay Count–Min Space Saving) which, to the best of our knowledge, is the world first message–passing parallel algorithm for mining time–faded heavy hitters. The algorithm is a parallel version of the recently published FDCMSS (Forward Decay Count–Min Space Saving) sequential algorithm. We formally prove its correctness by showing that the underlying data structure, a sketch augmented with a Space Saving stream summary holding exactly two counters, is mergeable. Whilst mergeability of traditional sketches derives immediately from theory, we show that, instead, merging our augmented sketch is non trivial. Nonetheless, the resulting parallel algorithm is fast and simple to implement. The very large volumes of modern datasets in the context of Big Data present new challenges that current sequential algorithms can not cope with; on the contrary, parallel computing enables near real time processing of very large datasets, which are growing at an unprecedented scale. Our algorithm’s implementation, taking advantage of the MPI (Message Passing Interface) library, is portable, reliable and provides cutting–edge performance. Extensive experimental results confirm that PFDCMSS retains the extreme accuracy and error bound provided by FDCMSS whilst providing excellent parallel scalability. Our contributions are three-fold: (i) we prove the non trivial mergeability of the augmented sketch used in the FDCMSS algorithm; (ii) we derive PFDCMSS, a novel message–passing parallel algorithm; (iii) we experimentally prove that PFDCMSS is extremely accurate and scalable, allowing near real time processing of large datasets. The result supports both casual users and seasoned, professional scientists working on expert and intelligent systems."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | @cite_12 addressed the FS scaling problem using an asynchronous search approach, given that synchronous search, as commonly performed, can lead to efficiency losses due to the inactivity of some processors waiting for other processors to end their tasks. In their tests, they first obtained an initial reduction using a mutual information (MI) @cite_9 filter and then evaluated subsets using a random forest (RF) @cite_14 classifier. However, as stated by those authors, any other approach could be used for subset evaluation. | {
"cite_N": [
"@cite_9",
"@cite_14",
"@cite_12"
],
"mid": [
"2952294252",
"2587322366",
"2531879448"
],
"abstract": [
"This paper describes a novel approach of packing sparse convolutional neural networks for their efficient systolic array implementations. By combining subsets of columns in the original filter matrix associated with a convolutional layer, we increase the utilization efficiency of the systolic array substantially (e.g., 4x) due to the increased density of nonzeros in the resulting packed filter matrix. In combining columns, for each row, all filter weights but one with the largest magnitude are pruned. We retrain the remaining weights to preserve high accuracy. We demonstrate that in mitigating data privacy concerns the retraining can be accomplished with only fractions of the original dataset (e.g., 10 for CIFAR-10). We study the effectiveness of this joint optimization for both high utilization and classification accuracy with ASIC and FPGA designs based on efficient bit-serial implementations of multiplier-accumulators. We present analysis and empirical evidence on the superior performance of our column combining approach against prior arts under metrics such as energy efficiency (3x) and inference latency (12x).",
"Reducing the dimensionality of datasets is a fundamental step in the task of building a classification model. Feature selection is the process of selecting a smaller subset of features from the original one in order to enhance the performance of the classification model. The problem is known to be NP-hard, and despite the existence of several algorithms there is not one that outperforms the others in all scenarios. Due to the complexity of the problem usually feature selection algorithms have to compromise the quality of their solutions in order to execute in a practicable amount of time. Parallel computing techniques emerge as a potential solution to tackle this problem. There are several approaches that already execute feature selection in parallel resorting to synchronous models. These are preferred due to their simplicity and capability to use with any feature selection algorithm. However, synchronous models implement pausing points during the execution flow, which decrease the parallel performance. In this paper, we discuss the challenges of executing feature selection algorithms in parallel using asynchronous models, and present a feature selection algorithm that favours these models. Furthermore, we present two strategies for an asynchronous parallel execution not only of our algorithm but of any other feature selection approach. The first strategy solves the problem using the distributed memory paradigm, while the second exploits the use of shared memory. We evaluate the parallel performance of our strategies using up to 32 cores. The results show near linear speedups for both strategies, with the shared memory strategy outperforming the distributed one. Additionally, we provide an example of adapting our strategies to execute the Sequential forward Search asynchronously. We further test this version versus a synchronous one. Our results revealed that, by using an asynchronous strategy, we are able to save an average of 7.5 of the execution time.",
"Numerous researchers have studied the contention that arises among tasks running in parallel on a multicore processor. Most of those studies seek to derive a tight and sound upper-bound for the worst-case delay with which a processor resource may serve an incoming request, when its access is arbitrated using time-predictable policies such as round-robin or FIFO. We call this value upper-bound delay ( @math ). Deriving trustworthy @math statically is possible when sufficient public information exists on the timing latency incurred on access to the resource of interest. Unfortunately however, that is rarely granted for commercial-of-the-shelf (COTS) processors. Therefore, the users resort to measurement observations on the target processor and thus compute a “measured” @math . However, using @math to compute worst-case execution time values for programs running on COTS multicore processors requires qualification on the soundness of the result. In this paper, we present a measurement-based methodology to derive a @math under round-robin (RoRo) and first-in-first-out (FIFO) arbitration, which accurately approximates @math from above, without needing latency information from the hardware provider. Experimental results, obtained on multiple processor configurations, demonstrate the robustness of the proposed methodology."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | @cite_23 used the MapReduce model to implement a wrapper-based evolutionary search FS method. The dataset was split by instances and the FS method was applied to each resulting subset. Simple majority voting was used as a reduction step for the selected features and the final subset of feature was selected according to a user-defined threshold. All tests were carried out using the EPSILON dataset, which we also use here (see ). | {
"cite_N": [
"@cite_23"
],
"mid": [
"1952835952"
],
"abstract": [
"Nowadays, many disciplines have to deal with big datasets that additionally involve a high number of features. Feature selection methods aim at eliminating noisy, redundant, or irrelevant features that may deteriorate the classification performance. However, traditional methods lack enough scalability to cope with datasets of millions of instances and extract successful results in a delimited time. This paper presents a feature selection algorithm based on evolutionary computation that uses the MapReduce paradigm to obtain subsets of features from big datasets. The algorithm decomposes the original dataset in blocks of instances to learn from them in the map phase; then, the reduce phase merges the obtained partial results into a final vector of feature weights, which allows a flexible application of the feature selection procedure using a threshold to determine the selected subset of features. The feature selection method is evaluated by using three well-known classifiers (SVM, Logistic Regression, and Naive Bayes) implemented within the Spark framework to address big data problems. In the experiments, datasets up to 67 millions of instances and up to 2000 attributes have been managed, showing that this is a suitable framework to perform evolutionary feature selection, improving both the classification accuracy and its runtime when dealing with big data problems."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | Bol 'on- @cite_39 proposed a framework to deal with high dimensionality data by first optionally ranking features using a FS filter, then partitioning vertically by dividing the data according to features (columns) rather than, as commonly done, according to instances (rows). After partitioning, another FS filter is applied to each partition, and finally, a merging procedure guided by a classifier obtains a single set of features. The authors experiment with five commonly used FS filters for the partitions, namely, CFS @cite_33 , Consistency @cite_0 , INTERACT @cite_30 , Information Gain @cite_6 and ReliefF @cite_13 , and with four classifiers for the final merging, namely, C4.5 @cite_3 , Naive Bayes @cite_40 , @math -Nearest Neighbors @cite_22 and SVM @cite_35 , show that their own approach significantly reduces execution times while maintaining and, in some cases, even improving accuracy. | {
"cite_N": [
"@cite_30",
"@cite_35",
"@cite_33",
"@cite_22",
"@cite_6",
"@cite_39",
"@cite_0",
"@cite_3",
"@cite_40",
"@cite_13"
],
"mid": [
"1968752118",
"2773012335",
"2952294252",
"1994576156",
"2342423902",
"2963727650",
"2789983685",
"2802962644",
"1495061682",
"1992793057"
],
"abstract": [
"In order to encode the class correlation and class specific information in image representation, we propose a new local feature learning approach named Deep Discriminative and Shareable Feature Learning (DDSFL). DDSFL aims to hierarchically learn feature transformation filter banks to transform raw pixel image patches to features. The learned filter banks are expected to (1) encode common visual patterns of a flexible number of categories; (2) encode discriminative information; and (3) hierarchically extract patterns at different visual levels. Particularly, in each single layer of DDSFL, shareable filters are jointly learned for classes which share the similar patterns. Discriminative power of the filters is achieved by enforcing the features from the same category to be close, while features from different categories to be far away from each other. Furthermore, we also propose two exemplar selection methods to iteratively select training data for more efficient and effective learning. Based on the experimental results, DDSFL can achieve very promising performance, and it also shows great complementary effect to the state-of-the-art Caffe features. HighlightsWe propose to encode shareable and discriminative information in feature learning.Two exemplar selection methods are proposed to select effective training data.We build a hierarchical learning scheme to capture multiple visual level information.Our DDSFL outperforms most of the existing features.DDSFL features show great complementary effect to Caffe features.",
"In this paper, we focus on improving the proposal classification stage in the object detection task and present implicit negative sub-categorization and sink diversion to lift the performance by strengthening loss function in this stage. First, based on the observation that the “background” class is generally very diverse and thus challenging to be handled as a single indiscriminative class in existing state-of-the-art methods, we propose to divide the background category into multiple implicit sub-categories to explicitly differentiate diverse patterns within it. Second, since the ground truth class inevitably has low-value probability scores for certain images, we propose to add a “sink” class and divert the probabilities of wrong classes to this class when necessary, such that the ground truth label will still have a higher probability than other wrong classes even though it has low probability output. Additionally, we propose to use dilated convolution, which is widely used in the semantic segmentation task, for efficient and valuable context information extraction. Extensive experiments on PASCAL VOC 2007 and 2012 data sets show that our proposed methods based on faster R-CNN implementation can achieve state-of-the-art mAPs, i.e., 84.1 , 82.6 , respectively, and obtain 2.5 improvement on ILSVRC DET compared with that of ResNet.",
"This paper describes a novel approach of packing sparse convolutional neural networks for their efficient systolic array implementations. By combining subsets of columns in the original filter matrix associated with a convolutional layer, we increase the utilization efficiency of the systolic array substantially (e.g., 4x) due to the increased density of nonzeros in the resulting packed filter matrix. In combining columns, for each row, all filter weights but one with the largest magnitude are pruned. We retrain the remaining weights to preserve high accuracy. We demonstrate that in mitigating data privacy concerns the retraining can be accomplished with only fractions of the original dataset (e.g., 10 for CIFAR-10). We study the effectiveness of this joint optimization for both high utilization and classification accuracy with ASIC and FPGA designs based on efficient bit-serial implementations of multiplier-accumulators. We present analysis and empirical evidence on the superior performance of our column combining approach against prior arts under metrics such as energy efficiency (3x) and inference latency (12x).",
"A major challenge for collaborative filtering (CF) techniques in recommender systems is the data sparsity that is caused by missing and noisy ratings. This problem is even more serious for CF domains where the ratings are expressed numerically, e.g. as 5-star grades. We assume the 5-star ratings are unordered bins instead of ordinal relative preferences. We observe that, while we may lack the information in numerical ratings, we sometimes have additional auxiliary data in the form of binary ratings. This is especially true given that users can easily express themselves with their preferences expressed as likes or dislikes for items. In this paper, we explore how to use these binary auxiliary preference data to help reduce the impact of data sparsity for CF domains expressed in numerical ratings. We solve this problem by transferring the rating knowledge from some auxiliary data source in binary form (that is, likes or dislikes), to a target numerical rating matrix. In particular, our solution is to model both the numerical ratings and ratings expressed as like or dislike in a principled way. We present a novel framework of Transfer by Collective Factorization (TCF), in which we construct a shared latent space collectively and learn the data-dependent effect separately. A major advantage of the TCF approach over the previous bilinear method of collective matrix factorization is that we are able to capture the data-dependent effect when sharing the data-independent knowledge. This allows us to increase the overall quality of knowledge transfer. We present extensive experimental results to demonstrate the effectiveness of TCF at various sparsity levels, and show improvements of our approach as compared to several state-of-the-art methods.",
"Local descriptors are mostly extracted over the raw intensity values of the image and requires the analysis of patterns in local pixel neighborhoods. The raw images don't provide much local relationship information required for the local descriptors. To resolve this problem, a new method of local image feature description is proposed in this paper based on the patterns extracted from the multiple filtered images. The whole process is structured as follows: First, the images are filtered with the bag of filters (BoF) and then local binary pattern (LBP) is computed over each filtered image and finally concatenated to find out the single BoF-LBP descriptor. The local information enriches in the filtered images improves the discriminative ability of the final descriptor. Content based image retrieval (CBIR) experiments are performed to observe the effectiveness of the proposed approach and compared with the recent state-of-the-art methods. The experiments are conducted over four benchmark databases having natural and texture images such as Corel-1k, Corel-10k, MIT-VisTex and STex-512S. The experimental results confirm that the introduced approach is able to improve the retrieval performance of the CBIR system.",
"Recent work has made significant progress in improving spatial resolution for pixelwise labeling with Fully Convolutional Network (FCN) framework by employing Dilated Atrous convolution, utilizing multi-scale features and refining boundaries. In this paper, we explore the impact of global contextual information in semantic segmentation by introducing the Context Encoding Module, which captures the semantic context of scenes and selectively highlights class-dependent featuremaps. The proposed Context Encoding Module significantly improves semantic segmentation results with only marginal extra computation cost over FCN. Our approach has achieved new state-of-the-art results 51.7 mIoU on PASCAL-Context, 85.9 mIoU on PASCAL VOC 2012. Our single model achieves a final score of 0.5567 on ADE20K test set, which surpasses the winning entry of COCO-Place Challenge 2017. In addition, we also explore how the Context Encoding Module can improve the feature representation of relatively shallow networks for the image classification on CIFAR-10 dataset. Our 14 layer network has achieved an error rate of 3.45 , which is comparable with state-of-the-art approaches with over 10A— more layers. The source code for the complete system are publicly available1.",
"Recent work has made significant progress in improving spatial resolution for pixelwise labeling with Fully Convolutional Network (FCN) framework by employing Dilated Atrous convolution, utilizing multi-scale features and refining boundaries. In this paper, we explore the impact of global contextual information in semantic segmentation by introducing the Context Encoding Module, which captures the semantic context of scenes and selectively highlights class-dependent featuremaps. The proposed Context Encoding Module significantly improves semantic segmentation results with only marginal extra computation cost over FCN. Our approach has achieved new state-of-the-art results 51.7 mIoU on PASCAL-Context, 85.9 mIoU on PASCAL VOC 2012. Our single model achieves a final score of 0.5567 on ADE20K test set, which surpass the winning entry of COCO-Place Challenge in 2017. In addition, we also explore how the Context Encoding Module can improve the feature representation of relatively shallow networks for the image classification on CIFAR-10 dataset. Our 14 layer network has achieved an error rate of 3.45 , which is comparable with state-of-the-art approaches with over 10 times more layers. The source code for the complete system are publicly available.",
"With the development of Fully Convolutional Neural Network (FCN), there have been progressive advances in the field of semantic segmentation in recent years. The FCN-based solutions are able to summarize features across training images and generate matching templates for the desired object classes, yet they overlook intra-class difference (ICD) among multiple instances in the same class. In this work, we present a novel fine-to-coarse learning (FCL) procedure, which first guides the network with designed 'finer' sub-class labels, whose decisions are mapped to the original 'coarse' object category through end-to-end learning. A sub-class labeling strategy is designed with unsupervised clustering upon deep convolutional features, and the proposed FCL procedure enables a balance between the fine-scale (i.e. sub-class) and the coarse-scale (i.e. class) knowledge. We conduct extensive experiments on several popular datasets, including PASCAL VOC, Context, Person-Part and NYUDepth-v2 to demonstrate the advantage of learning finer sub-classes and the potential to guide the learning of deep networks with unsupervised clustering.",
"A central problem in machine learning is identifying a representative set of features from which to construct a classification model for a particular task. This thesis addresses the problem of feature selection for machine learning through a correlation based approach. The central hypothesis is that good feature sets contain features that are highly correlated with the class, yet uncorrelated with each other. A feature evaluation formula, based on ideas from test theory, provides an operational definition of this hypothesis. CFS (Correlation based Feature Selection) is an algorithm that couples this evaluation formula with an appropriate correlation measure and a heuristic search strategy. CFS was evaluated by experiments on artificial and natural datasets. Three machine learning algorithms were used: C4.5 (a decision tree learner), IB1 (an instance based learner), and naive Bayes. Experiments on artificial datasets showed that CFS quickly identifies and screens irrelevant, redundant, and noisy features, and identifies relevant features as long as their relevance does not strongly depend on other features. On natural domains, CFS typically eliminated well over half the features. In most cases, classification accuracy using the reduced feature set equaled or bettered accuracy using the complete feature set. Feature selection degraded machine learning performance in cases where some features were eliminated which were highly predictive of very small areas of the instance space. Further experiments compared CFS with a wrapper—a well known approach to feature selection that employs the target learning algorithm to evaluate feature sets. In many cases CFS gave comparable results to the wrapper, and in general, outperformed the wrapper on small datasets. CFS executes many times faster than the wrapper, which allows it to scale to larger datasets. Two methods of extending CFS to handle feature interaction are presented and experimentally evaluated. The first considers pairs of features and the second incorporates iii feature weights calculated by the RELIEF algorithm. Experiments on artificial domains showed that both methods were able to identify interacting features. On natural domains, the pairwise method gave more reliable results than using weights provided by RELIEF.",
"We used both inner and outer face cues.External cues improve classification performance for gender recognition.FIS framework improves classification results when combined with SVM.Unconstrained databases provide better results than that of constrained databases.We obtained 93.35 accuracy on Groups LFW cross-database test. In this paper, we propose a novel gender recognition framework based on a fuzzy inference system (FIS). Our main objective is to study the gain brought by FIS in presence of various visual sensors (e.g., hair, mustache, inner face). We use inner and outer facial features to extract input variables. First, we define the fuzzy statements and then we generate a knowledge base composed of a set of rules over the linguistic variables including hair volume, mustache and a vision-sensor. Hair volume and mustache information are obtained from Part Labels subset of Labeled Faces in the Wild (LFW) database and vision-sensor is obtained from a pixel-intensity based SVM+RBF classifier trained on different databases including Feret, Groups and GENKI-4K. Cross-database test experiments on LFW database showed that the proposed method provides better accuracy than optimized SVM+RBF only classification. We also showed that FIS increases the inter-class variability by decreasing false negatives (FN) and false positives (FP) using expert knowledge. Our experimental results yield an average accuracy of 93.35 using Groups LFW test, while the SVM performance baseline yields 91.25 accuracy."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | @cite_24 described a distributed parallel FS method based on a variance preservation criterion using the proprietary software SAS High-Performance Analytics. http: www.sas.com en_us software high-performance-analytics.html One remarkable characteristic of the method is its support not only for supervised FS, but also for unsupervised FS where no label information is available. Their experiments were carried out with datasets with both high dimensionality and a high number of instances. | {
"cite_N": [
"@cite_24"
],
"mid": [
"2386452692"
],
"abstract": [
"Lasso regression is a widely used technique in data mining for model selection and feature extraction. In many applications, it remains challenging to apply the regression model to large-scale problems that have massive data samples with high-dimensional features. One popular and promising strategy is to solve the Lasso problem in parallel. Parallel solvers run multiple cores in parallel on a shared memory system to speedup the computation, while the practical usage is limited by the huge dimension in the feature space. Screening is a promising method to solve the problem of high dimensionality by discarding the inactive features and removing them from optimization. However, when integrating screening methods with parallel solvers, most of solvers cannot guarantee the convergence on the reduced feature matrix. In this paper, we propose a novel parallel framework by parallelizing screening methods and integrating it with our proposed parallel solver. We propose two parallel screening algorithms: Parallel Strong Rule (PSR) and Parallel Dual Polytope Projection (PDPP). For the parallel solver, we proposed an Asynchronous Grouped Coordinate Descent method (AGCD) to optimize the regression problem in parallel on the reduced feature matrix. AGCD is based on a grouped selection strategy to select the coordinate that has the maximum descent for the objective function in a group of candidates. Empirical studies on the real-world datasets demonstrate that the proposed parallel framework has a superior performance compared to the state-of-the-art parallel solvers."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | Ram 'irez- @cite_28 described scalable versions of the popular mRMR @cite_9 FS filter that included a distributed version using Spark. The authors showed that their version that leveraged the power of a cluster of computers could perform much faster than the original and processed much larger datasets. | {
"cite_N": [
"@cite_28",
"@cite_9"
],
"mid": [
"2200550062",
"2475596014"
],
"abstract": [
"There has been an increased growth in a number of applications that naturally generate large volumes of uncertain data. By the advent of such applications, the support of advanced analysis queries such as the skyline and its variant operators for big uncertain data has become important. In this paper, we propose the effective parallel algorithms using MapReduce to process the probabilistic skyline queries for uncertain data modeled by both discrete and continuous models. We present three filtering methods to identify probabilistic non-skyline objects in advance. We next develop a single MapReduce phase algorithm PS-QP-MR by utilizing space partitioning based on a variant of quadtrees to distribute the instances of objects effectively and the enhanced algorithm PS-QPF-MR by applying the three filtering methods additionally. We also propose the workload balancing technique to balance the workload of reduce functions based on the number of machines available. Finally, we present the brute-force algorithms PS-BR-MR and PS-BRF-MR with partitioning randomly and applying the filtering methods. In our experiments, we demonstrate the efficiency and scalability of PS-QPF-MR compared to the other algorithms.",
"With the advent of large-scale problems, feature selection has become a fundamental preprocessing step to reduce input dimensionality. The minimum-redundancy-maximum-relevance (mRMR) selector is considered one of the most relevant methods for dimensionality reduction due to its high accuracy. However, it is a computationally expensive technique, sharply affected by the number of features. This paper presents fast-mRMR, an extension of mRMR, which tries to overcome this computational burden. Associated with fast-mRMR, we include a package with three implementations of this algorithm in several platforms, namely, CPU for sequential execution, GPU (graphics processing units) for parallel computing, and Apache Spark for distributed computing using big data technologies."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | In a previous work @cite_36 , using the Spark computing model we designed a distributed version of the ReliefF @cite_13 filter, called DiReliefF. In testing using datasets with large numbers of features and instances, it was much more efficient and scalable than the original filter. | {
"cite_N": [
"@cite_36",
"@cite_13"
],
"mid": [
"2785172785",
"1560107318"
],
"abstract": [
"Feature selection (FS) is a key research area in the machine learning and data mining fields; removing irrelevant and redundant features usually helps to reduce the effort required to process a dataset while maintaining or even improving the processing algorithm’s accuracy. However, traditional algorithms designed for executing on a single machine lack scalability to deal with the increasing amount of data that have become available in the current Big Data era. ReliefF is one of the most important algorithms successfully implemented in many FS applications. In this paper, we present a completely redesigned distributed version of the popular ReliefF algorithm based on the novel Spark cluster computing model that we have called DiReliefF. The effectiveness of our proposal is tested on four publicly available datasets, all of them with a large number of instances and two of them with also a large number of features. Subsets of these datasets were also used to compare the results to a non-distributed implementation of the algorithm. The results show that the non-distributed implementation is unable to handle such large volumes of data without specialized hardware, while our design can process them in a scalable way with much better processing times and memory usage.",
"Algorithms for feature selection fall into two broad categories: wrappers that use the learning algorithm itself to evaluate the usefulness of features and filters that evaluate features according to heuristics based on general characteristics of the data. For application to large databases, filters have proven to be more practical than wrappers because they are much faster. However, most existing filter algorithms only work with discrete classification problems. This paper describes a fast, correlation-based filter algorithm that can be applied to continuous and discrete problems. The algorithm often outperforms the well-known ReliefF attribute estimator when used as a preprocessing step for naive Bayes, instance-based learning, decision trees, locally weighted regression, and model trees. It performs more feature selection than ReliefF does—reducing the data dimensionality by fifty percent in most cases. Also, decision and model trees built from the preprocessed data are often significantly smaller."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | Finally, Eiras- @cite_32 , using four distributed FS algorithms, three of them filters, namely, InfoGain @cite_6 , ReliefF @cite_13 and the CFS @cite_33 , reduce execution times with respect to the original versions. However, in the CFS case, the version of those authors focuses on regression problems where all the features, including the class label, are numerical, with correlations calculated using the Pearson coefficient. A completely different approach is required to design a parallel version for classification problems where correlations are based on the information theory. | {
"cite_N": [
"@cite_33",
"@cite_13",
"@cite_32",
"@cite_6"
],
"mid": [
"1495061682",
"1994576156",
"2000292092",
"2142057089"
],
"abstract": [
"A central problem in machine learning is identifying a representative set of features from which to construct a classification model for a particular task. This thesis addresses the problem of feature selection for machine learning through a correlation based approach. The central hypothesis is that good feature sets contain features that are highly correlated with the class, yet uncorrelated with each other. A feature evaluation formula, based on ideas from test theory, provides an operational definition of this hypothesis. CFS (Correlation based Feature Selection) is an algorithm that couples this evaluation formula with an appropriate correlation measure and a heuristic search strategy. CFS was evaluated by experiments on artificial and natural datasets. Three machine learning algorithms were used: C4.5 (a decision tree learner), IB1 (an instance based learner), and naive Bayes. Experiments on artificial datasets showed that CFS quickly identifies and screens irrelevant, redundant, and noisy features, and identifies relevant features as long as their relevance does not strongly depend on other features. On natural domains, CFS typically eliminated well over half the features. In most cases, classification accuracy using the reduced feature set equaled or bettered accuracy using the complete feature set. Feature selection degraded machine learning performance in cases where some features were eliminated which were highly predictive of very small areas of the instance space. Further experiments compared CFS with a wrapper—a well known approach to feature selection that employs the target learning algorithm to evaluate feature sets. In many cases CFS gave comparable results to the wrapper, and in general, outperformed the wrapper on small datasets. CFS executes many times faster than the wrapper, which allows it to scale to larger datasets. Two methods of extending CFS to handle feature interaction are presented and experimentally evaluated. The first considers pairs of features and the second incorporates iii feature weights calculated by the RELIEF algorithm. Experiments on artificial domains showed that both methods were able to identify interacting features. On natural domains, the pairwise method gave more reliable results than using weights provided by RELIEF.",
"A major challenge for collaborative filtering (CF) techniques in recommender systems is the data sparsity that is caused by missing and noisy ratings. This problem is even more serious for CF domains where the ratings are expressed numerically, e.g. as 5-star grades. We assume the 5-star ratings are unordered bins instead of ordinal relative preferences. We observe that, while we may lack the information in numerical ratings, we sometimes have additional auxiliary data in the form of binary ratings. This is especially true given that users can easily express themselves with their preferences expressed as likes or dislikes for items. In this paper, we explore how to use these binary auxiliary preference data to help reduce the impact of data sparsity for CF domains expressed in numerical ratings. We solve this problem by transferring the rating knowledge from some auxiliary data source in binary form (that is, likes or dislikes), to a target numerical rating matrix. In particular, our solution is to model both the numerical ratings and ratings expressed as like or dislike in a principled way. We present a novel framework of Transfer by Collective Factorization (TCF), in which we construct a shared latent space collectively and learn the data-dependent effect separately. A major advantage of the TCF approach over the previous bilinear method of collective matrix factorization is that we are able to capture the data-dependent effect when sharing the data-independent knowledge. This allows us to increase the overall quality of knowledge transfer. We present extensive experimental results to demonstrate the effectiveness of TCF at various sparsity levels, and show improvements of our approach as compared to several state-of-the-art methods.",
"Many machine learning and pattern classification methods have been applied to the diagnosis of Alzheimer’s disease (AD) and its prodromal stage, i.e., mild cognitive impairment (MCI). Recently, rather than predicting categorical variables as in classification, several pattern regression methods have also been used to estimate continuous clinical variables from brain images. However, most existing regression methods focus on estimating multiple clinical variables separately and thus cannot utilize the intrinsic useful correlation information among different clinical variables. On the other hand, in those regression methods, only a single modality of data (usually only the structural MRI) is often used, without considering the complementary information that can be provided by different modalities. In this paper, we propose a general methodology, namely Multi-Modal Multi-Task (M3T) learning, to jointly predict multiple variables from multi-modal data. Here, the variables include not only the clinical variables used for regression but also the categorical variable used for classification, with different tasks corresponding to prediction of different variables. Specifically, our method contains two key components, i.e., (1) a multi-task feature selection which selects the common subset of relevant features for multiple variables from each modality, and (2) a multi-modal support vector machine which fuses the above-selected features from all modalities to predict multiple (regression and classification) variables. To validate our method, we perform two sets of experiments on ADNI baseline MRI, FDG-PET, and cerebrospinal fluid (CSF) data from 45 AD patients, 91 MCI patients, and 50 healthy controls (HC). In the first set of experiments, we estimate two clinical variables such as Mini Mental State Examination (MMSE) and Alzheimer’s Disease Assessment Scale - Cognitive Subscale (ADAS-Cog), as well as one categorical variable (with value of ‘AD’, ‘MCI’ or ‘HC’), from the baseline MRI, FDG-PET, and CSF data. In the second set of experiments, we predict the 2-year changes of MMSE and ADAS-Cog scores and also the conversion of MCI to AD from the baseline MRI, FDG-PET, and CSF data. The results on both sets of experiments demonstrate that our proposed M3T learning scheme can achieve better performance on both regression and classification tasks than the conventional learning methods.",
"Recommender problems with large and dynamic item pools are ubiquitous in web applications like content optimization, online advertising and web search. Despite the availability of rich item meta-data, excess heterogeneity at the item level often requires inclusion of item-specific \"factors\" (or weights) in the model. However, since estimating item factors is computationally intensive, it poses a challenge for time-sensitive recommender problems where it is important to rapidly learn factors for new items (e.g., news articles, event updates, tweets) in an online fashion. In this paper, we propose a novel method called FOBFM (Fast Online Bilinear Factor Model) to learn item-specific factors quickly through online regression. The online regression for each item can be performed independently and hence the procedure is fast, scalable and easily parallelizable. However, the convergence of these independent regressions can be slow due to high dimensionality. The central idea of our approach is to use a large amount of historical data to initialize the online models based on offline features and learn linear projections that can effectively reduce the dimensionality. We estimate the rank of our linear projections by taking recourse to online model selection based on optimizing predictive likelihood. Through extensive experiments, we show that our method significantly and uniformly outperforms other competitive methods and obtains relative lifts that are in the range of 10-15 in terms of predictive log-likelihood, 200-300 for a rank correlation metric on a proprietary My Yahoo! dataset; it obtains 9 reduction in root mean squared error over the previously best method on a benchmark MovieLens dataset using a time-based train test data split."
]
} |
1901.11286 | 2899274556 | Abstract Feature selection (FS) is a key preprocessing step in data mining. CFS (Correlation-Based Feature Selection) is an FS algorithm that has been successfully applied to classification problems in many domains. We describe Distributed CFS (DiCFS) as a completely redesigned, scalable, parallel and distributed version of the CFS algorithm, capable of dealing with the large volumes of data typical of big data applications. Two versions of the algorithm were implemented and compared using the Apache Spark cluster computing model, currently gaining popularity due to its much faster processing times than Hadoop’s MapReduce model. We tested our algorithms on four publicly available datasets, each consisting of a large number of instances and two also consisting of a large number of features. The results show that our algorithms were superior in terms of both time-efficiency and scalability. In leveraging a computer cluster, they were able to handle larger datasets than the non-distributed WEKA version while maintaining the quality of the results, i.e., exactly the same features were returned by our algorithms when compared to the original algorithm available in WEKA. | The approach described here can be categorized as a approach that builds on works described elsewhere @cite_28 , @cite_36 , @cite_32 . The fact that their focus was not only on designing an efficient and scalable FS algorithm, but also on preserving the original behaviour (and obtaining the same final results) of traditional filters, means that research focused on those filters is also valid for adapted versions. Another important issue in relation to filters is that, since they are generally more efficient than wrappers, they are often the only feasible option due to the abundance of data. It is worth mentioning that scalable filters could feasibly be included in any of the methods mentioned in the and categories, where an initial filtering step is implemented to improve performance. | {
"cite_N": [
"@cite_28",
"@cite_32",
"@cite_36"
],
"mid": [
"2952294252",
"2063228777",
"2101207848"
],
"abstract": [
"This paper describes a novel approach of packing sparse convolutional neural networks for their efficient systolic array implementations. By combining subsets of columns in the original filter matrix associated with a convolutional layer, we increase the utilization efficiency of the systolic array substantially (e.g., 4x) due to the increased density of nonzeros in the resulting packed filter matrix. In combining columns, for each row, all filter weights but one with the largest magnitude are pruned. We retrain the remaining weights to preserve high accuracy. We demonstrate that in mitigating data privacy concerns the retraining can be accomplished with only fractions of the original dataset (e.g., 10 for CIFAR-10). We study the effectiveness of this joint optimization for both high utilization and classification accuracy with ASIC and FPGA designs based on efficient bit-serial implementations of multiplier-accumulators. We present analysis and empirical evidence on the superior performance of our column combining approach against prior arts under metrics such as energy efficiency (3x) and inference latency (12x).",
"The traditional collaborative filtering approaches have been shown to suffer from two fundamental problems: data sparsity and difficulty in scalability. To address these problems, we present a novel scalable item-based collaborative filtering method by using incremental update and local link prediction. By subdividing the computations and analyzing the factors in different cases of item-to-item similarity, we design the incremental update strategies in item-based CF, which can make the recommender system more efficient and scalable. Based on the transitive structure of item similarity graph, we use the local link prediction method to find implicit candidates to alleviate the lack of neighbors in predictions and recommendations caused by the sparsity of data. The experiment results validate that our algorithm can improve the performance of traditional CF, and can increase the efficiency in recommendations.",
"We consider the problem of pipelined filters, where a continuous stream of tuples is processed by a set of commutative filters. Pipelined filters are common in stream applications and capture a large class of multiway stream joins. We focus on the problem of ordering the filters adaptively to minimize processing cost in an environment where stream and filter characteristics vary unpredictably over time. Our core algorithm, A-Greedy (for Adaptive Greedy), has strong theoretical guarantees: If stream and filter characteristics were to stabilize, A-Greedy would converge to an ordering within a small constant factor of optimal. (In experiments A-Greedy usually converges to the optimal ordering.) One very important feature of A-Greedy is that it monitors and responds to selectivities that are correlated across filters (i.e., that are nonindependent), which provides the strong quality guarantee but incurs run-time overhead. We identify a three-way tradeoff among provable convergence to good orderings, run-time overhead, and speed of adaptivity. We develop a suite of variants of A-Greedy that lie at different points on this tradeoff spectrum. We have implemented all our algorithms in the STREAM prototype Data Stream Management System and a thorough performance evaluation is presented."
]
} |
1901.11524 | 2950533301 | We establish geometric and topological properties of the space of value functions in finite state-action Markov decision processes. Our main contribution is the characterization of the nature of its shape: a general polytope (, 2010). To demonstrate this result, we exhibit several properties of the structural relationship between policies and value functions including the line theorem, which shows that the value functions of policies constrained on all but one state describe a line segment. Finally, we use this novel perspective to introduce visualizations to enhance the understanding of the dynamics of reinforcement learning algorithms. | The dual formulation consists of maximizing the expected return for a given initial state distribution, as a function of the discounted state action visit frequency distribution. Contrary to the primal form, any feasible discounted state action visit frequency distribution maps to an policy @cite_9 . | {
"cite_N": [
"@cite_9"
],
"mid": [
"2149418961"
],
"abstract": [
"We investigate the dual approach to dynamic programming and reinforcement learning, based on maintaining an explicit representation of stationary distributions as opposed to value functions. A significant advantage of the dual approach is that it allows one to exploit well developed techniques for representing, approximating and estimating probability distributions, without running the risks associated with divergent value function estimation. A second advantage is that some distinct algorithms for the average reward and discounted reward case in the primal become unified under the dual. In this paper, we present a modified dual of the standard linear program that guarantees a globally normalized state visit distribution is obtained. With this reformulation, we then derive novel dual forms of dynamic programming, including policy evaluation, policy iteration and value iteration. Moreover, we derive dual formulations of temporal difference learning to obtain new forms of Sarsa and Q-learning. Finally, we scale these techniques up to large domains by introducing approximation, and develop new approximate off-policy learning algorithms that avoid the divergence problems associated with the primal approach. We show that the dual view yields a viable alternative to standard value function based techniques and opens new avenues for solving dynamic programming and reinforcement learning problems"
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | Chase and Kamara @cite_19 introduced the notion of graph encryption while they were presenting structured encryption as a generalization of searchable symmetric encryption (SSE) proposed by Song @cite_23 . They presented schemes for , and on labeled graph-structured data. In all of their proposed schemes, the graph was considered as an adjacency matrix and each entry was encrypted separately using symmetric key encryption. The main idea of their scheme, given a vertex and the corresponding key, the scheme could return adjacent vertices. However, complex query requires complex operation (like addition, subtraction, division etc.) on adjacent matrix which make the scheme unsuitable. | {
"cite_N": [
"@cite_19",
"@cite_23"
],
"mid": [
"2614426900",
"2063575624"
],
"abstract": [
"Driven by the growing security demands of data outsourcing applications in sustainable smart cities, encrypting clients’ data has been widely accepted by academia and industry. Data encryptions should be done at the client side before outsourcing, because clouds and edges are not trusted. Therefore, how to properly encrypt data in a way that the encrypted and remotely stored data can still be queried has become a challenging issue. Though keyword searches over encrypted textual data have been extensively studied, approaches for encrypting graph-structured data with support for answering graph queries are still lacking in the literature. In this paper, we specially investigate graph encryption method for an important graph query type, called top-k Nearest Keyword (kNK) searches. We design several indexes to store necessary information for answering queries and guarantee that private information about the graph such as vertex identifiers, keywords and edges are encrypted or excluded. Security and efficiency of our graph encryption scheme are demonstrated by theoretical proofs and experiments on real-world datasets, respectively.",
"We propose graph encryption schemes that efficiently support approximate shortest distance queries on large-scale encrypted graphs. Shortest distance queries are one of the most fundamental graph operations and have a wide range of applications. Using such graph encryption schemes, a client can outsource large-scale privacy-sensitive graphs to an untrusted server without losing the ability to query it. Other applications include encrypted graph databases and controlled disclosure systems. We propose GRECS (stands for GRaph EnCryption for approximate Shortest distance queries) which includes three oracle encryption schemes that are provably secure against any semi-honest server. Our first construction makes use of only symmetric-key operations, resulting in a computationally-efficient construction. Our second scheme makes use of somewhat-homomorphic encryption and is less computationally-efficient but achieves optimal communication complexity (i.e. uses a minimal amount of bandwidth). Finally, our third scheme is both computationally-efficient and achieves optimal communication complexity at the cost of a small amount of additional leakage. We implemented and evaluated the efficiency of our constructions experimentally. The experiments demonstrate that our schemes are efficient and can be applied to graphs that scale up to 1.6 million nodes and 11 million edges."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | A parallel secure computation framework has been designed and implemented by Nayak @cite_12 . This framework computes functions like histogram, PageRank, matrix factorization etc. To run this algorithms, introduced parallel programming paradigms to secure computation. The parallel and secure execution enables the algorithms to perform even for large datasets. However, they adopt Path-ORAM @cite_21 based techniques which is inefficient if the client has little computation power or the client doesn't uses very large size RAM. | {
"cite_N": [
"@cite_21",
"@cite_12"
],
"mid": [
"1608459536",
"2074833026"
],
"abstract": [
"We propose introducing modern parallel programming paradigms to secure computation, enabling their secure execution on large datasets. To address this challenge, we present Graph SC, a framework that (i) provides a programming paradigm that allows non-cryptography experts to write secure code, (ii) brings parallelism to such secure implementations, and (iii) meets the need for obliviousness, thereby not leaking any private information. Using Graph SC, developers can efficiently implement an oblivious version of graph-based algorithms (including sophisticated data mining and machine learning algorithms) that execute in parallel with minimal communication overhead. Importantly, our secure version of graph-based algorithms incurs a small logarithmic overhead in comparison with the non-secure parallel version. We build Graph SC and demonstrate, using several algorithms as examples, that secure computation can be brought into the realm of practicality for big data analysis. Our secure matrix factorization implementation can process 1 million ratings in 13 hours, which is a multiple order-of-magnitude improvement over the only other existing attempt, which requires 3 hours to process 16K ratings.",
"This paper proposes a compiler-based programming framework that automatically translates user-written structured grid code into scalable parallel implementation code for GPU-equipped clusters. To enable such automatic translations, we design a small set of declarative constructs that allow the user to express stencil computations in a portable and implicitly parallel manner. Our framework translates the user-written code into actual implementation code in CUDA for GPU acceleration and MPI for node-level parallelization with automatic optimizations such as computation and communication overlapping. We demonstrate the feasibility of such automatic translations by implementing several structured grid applications in our framework. Experimental results on the TSUBAME2.0 GPU-based supercomputer show that the performance is comparable as hand-written code and good strong and weak scalability up to 256 GPUs."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | Sketch-based approximate shortest distance queries over encrypted graph have been studied by Meng @cite_16 . In the pre-processing stage, the client computes the sketches for every vertex that is useful for efficient shortest distance query. Instead of encrypting the graph directly, they encrypted the pre-processed data. Thus, in their scheme, there is no chance of getting information about the original graph. | {
"cite_N": [
"@cite_16"
],
"mid": [
"2770638201"
],
"abstract": [
"Constrained shortest distance (CSD) querying is one of the fundamental graph query primitives, which finds the shortest distance from an origin to a destination in a graph with a constraint that the total cost does not exceed a given threshold. CSD querying has a wide range of applications, such as routing in telecommunications and transportation. With an increasing prevalence of cloud computing paradigm, graph owners desire to outsource their graphs to cloud servers. In order to protect sensitive information, these graphs are usually encrypted before being outsourced to the cloud. This, however, imposes a great challenge to CSD querying over encrypted graphs. Since performing constraint filtering is an intractable task, existing work mainly focuses on unconstrained shortest distance queries. CSD querying over encrypted graphs remains an open research problem. In this paper, we propose Connor , a novel graph encryption scheme that enables approximate CSD querying. Connor is built based on an efficient, tree-based ciphertext comparison protocol, and makes use of symmetric-key primitives and the somewhat homomorphic encryption, making it computationally efficient. Using Connor , a graph owner can first encrypt privacy-sensitive graphs and then outsource them to the cloud server, achieving the necessary privacy without losing the ability of querying. Extensive experiments with real-world data sets demonstrate the effectiveness and efficiency of the proposed graph encryption scheme."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | Shen @cite_1 introduced and studied cloud-based approximate in encrypted graphs which finds the shortest distance with a constraint such that the total cost does not exceed a given threshold. | {
"cite_N": [
"@cite_1"
],
"mid": [
"2770638201"
],
"abstract": [
"Constrained shortest distance (CSD) querying is one of the fundamental graph query primitives, which finds the shortest distance from an origin to a destination in a graph with a constraint that the total cost does not exceed a given threshold. CSD querying has a wide range of applications, such as routing in telecommunications and transportation. With an increasing prevalence of cloud computing paradigm, graph owners desire to outsource their graphs to cloud servers. In order to protect sensitive information, these graphs are usually encrypted before being outsourced to the cloud. This, however, imposes a great challenge to CSD querying over encrypted graphs. Since performing constraint filtering is an intractable task, existing work mainly focuses on unconstrained shortest distance queries. CSD querying over encrypted graphs remains an open research problem. In this paper, we propose Connor , a novel graph encryption scheme that enables approximate CSD querying. Connor is built based on an efficient, tree-based ciphertext comparison protocol, and makes use of symmetric-key primitives and the somewhat homomorphic encryption, making it computationally efficient. Using Connor , a graph owner can first encrypt privacy-sensitive graphs and then outsource them to the cloud server, achieving the necessary privacy without losing the ability of querying. Extensive experiments with real-world data sets demonstrate the effectiveness and efficiency of the proposed graph encryption scheme."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | Exact distance has been computed on dynamic encrypted graphs in @cite_9 . Similar to our paper, this paper uses a proxy to reduce client-side computation and information leakage to the cloud. In the scheme, adjacency lists are stored in an inverted index. However, in a single query, the scheme leaks all the nodes reachable from the queried vertex which is a lot of information about the graph. For example, if the graph is complete, it reveals the whole graph. | {
"cite_N": [
"@cite_9"
],
"mid": [
"2063575624"
],
"abstract": [
"We propose graph encryption schemes that efficiently support approximate shortest distance queries on large-scale encrypted graphs. Shortest distance queries are one of the most fundamental graph operations and have a wide range of applications. Using such graph encryption schemes, a client can outsource large-scale privacy-sensitive graphs to an untrusted server without losing the ability to query it. Other applications include encrypted graph databases and controlled disclosure systems. We propose GRECS (stands for GRaph EnCryption for approximate Shortest distance queries) which includes three oracle encryption schemes that are provably secure against any semi-honest server. Our first construction makes use of only symmetric-key operations, resulting in a computationally-efficient construction. Our second scheme makes use of somewhat-homomorphic encryption and is less computationally-efficient but achieves optimal communication complexity (i.e. uses a minimal amount of bandwidth). Finally, our third scheme is both computationally-efficient and achieves optimal communication complexity at the cost of a small amount of additional leakage. We implemented and evaluated the efficiency of our constructions experimentally. The experiments demonstrate that our schemes are efficient and can be applied to graphs that scale up to 1.6 million nodes and 11 million edges."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | A graph encryption scheme, that supports top- @math nearest keyword search queries, has been proposed by Liu @cite_13 . They have made an encrypted index using order preserving encryption for searching. Together with lightweight symmetric key encryption schemes, homomorphic encryption is used to compute on encrypted data. | {
"cite_N": [
"@cite_13"
],
"mid": [
"2063575624"
],
"abstract": [
"We propose graph encryption schemes that efficiently support approximate shortest distance queries on large-scale encrypted graphs. Shortest distance queries are one of the most fundamental graph operations and have a wide range of applications. Using such graph encryption schemes, a client can outsource large-scale privacy-sensitive graphs to an untrusted server without losing the ability to query it. Other applications include encrypted graph databases and controlled disclosure systems. We propose GRECS (stands for GRaph EnCryption for approximate Shortest distance queries) which includes three oracle encryption schemes that are provably secure against any semi-honest server. Our first construction makes use of only symmetric-key operations, resulting in a computationally-efficient construction. Our second scheme makes use of somewhat-homomorphic encryption and is less computationally-efficient but achieves optimal communication complexity (i.e. uses a minimal amount of bandwidth). Finally, our third scheme is both computationally-efficient and achieves optimal communication complexity at the cost of a small amount of additional leakage. We implemented and evaluated the efficiency of our constructions experimentally. The experiments demonstrate that our schemes are efficient and can be applied to graphs that scale up to 1.6 million nodes and 11 million edges."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | Besides, Zheng @cite_11 proposed link prediction in decentralized social network preserving the privacy. Their construction split the link score into private and public parts and applied sparse logistic regression to find links based on the content of the users. However, the graph data was not considered to be encrypted in the privacy preserving link prediction schemes. | {
"cite_N": [
"@cite_11"
],
"mid": [
"2293003800"
],
"abstract": [
"We consider the privacy-preserving link prediction problem in decentralized online social network (OSNs). We formulate the problem as a sparse logistic regression problem and solve it with a novel decentralized two-tier method using alternating direction method of multipliers (ADMM). This method enables end users to collaborate with their online service providers without jeopardizing their data privacy. The method also grants end users fine-grained privacy control to their personal data by supporting arbitrary public private data split. Using real-world data, we show that our method enjoys various advantages including high prediction accuracy, balanced workload, and limited communication overhead. Additionally, we demonstrate that our method copes well with link reconstruction attack."
]
} |
1901.11308 | 2914465018 | Link Prediction is an important and well-studied problem for social networks. Given a snapshot of a graph, the link prediction problem predicts which new interactions between members are most likely to occur in the near future. As networks grow in size, data owners are forced to store the data in remote cloud servers which reveals sensitive information about the network. The graphs are therefore stored in encrypted form. We study the link prediction problem on encrypted graphs. To the best of our knowledge, this secure link prediction problem has not been studied before. We use the number of common neighbors for prediction. We present three algorithms for the secure link prediction problem. We design prototypes of the schemes and formally prove their security. We execute our algorithms in real-life datasets. | In this paper, we outsource the graph in encrypted form. In most of the previous works, the schemes are designed to perform single specific query like neighbor query ( @cite_19 ), shortest distance query ( @cite_16 @cite_1 @cite_9 ), focused subgraph queries ( @cite_19 ) etc. So, either it is hard to get the information about the source graph ( @cite_16 , @cite_1 ), as they do not support basic queries, or leaks a lot of information for a single query ( @cite_9 ). One trivial approach is that taking different schemes and use all of them to support all types of required queries. In this paper, our target is to get as much information about the graph as possible whenever required with supporting the link prediction query and leak as little information as possible. To the best of our knowledge, the secure link prediction problem has not been studied before. We study issues on link prediction problem in encrypted outsourced data and give three possible solutions overcoming them. | {
"cite_N": [
"@cite_19",
"@cite_9",
"@cite_16",
"@cite_1"
],
"mid": [
"2614426900",
"2770638201",
"1993294649",
"2063575624"
],
"abstract": [
"Driven by the growing security demands of data outsourcing applications in sustainable smart cities, encrypting clients’ data has been widely accepted by academia and industry. Data encryptions should be done at the client side before outsourcing, because clouds and edges are not trusted. Therefore, how to properly encrypt data in a way that the encrypted and remotely stored data can still be queried has become a challenging issue. Though keyword searches over encrypted textual data have been extensively studied, approaches for encrypting graph-structured data with support for answering graph queries are still lacking in the literature. In this paper, we specially investigate graph encryption method for an important graph query type, called top-k Nearest Keyword (kNK) searches. We design several indexes to store necessary information for answering queries and guarantee that private information about the graph such as vertex identifiers, keywords and edges are encrypted or excluded. Security and efficiency of our graph encryption scheme are demonstrated by theoretical proofs and experiments on real-world datasets, respectively.",
"Constrained shortest distance (CSD) querying is one of the fundamental graph query primitives, which finds the shortest distance from an origin to a destination in a graph with a constraint that the total cost does not exceed a given threshold. CSD querying has a wide range of applications, such as routing in telecommunications and transportation. With an increasing prevalence of cloud computing paradigm, graph owners desire to outsource their graphs to cloud servers. In order to protect sensitive information, these graphs are usually encrypted before being outsourced to the cloud. This, however, imposes a great challenge to CSD querying over encrypted graphs. Since performing constraint filtering is an intractable task, existing work mainly focuses on unconstrained shortest distance queries. CSD querying over encrypted graphs remains an open research problem. In this paper, we propose Connor , a novel graph encryption scheme that enables approximate CSD querying. Connor is built based on an efficient, tree-based ciphertext comparison protocol, and makes use of symmetric-key primitives and the somewhat homomorphic encryption, making it computationally efficient. Using Connor , a graph owner can first encrypt privacy-sensitive graphs and then outsource them to the cloud server, achieving the necessary privacy without losing the ability of querying. Extensive experiments with real-world data sets demonstrate the effectiveness and efficiency of the proposed graph encryption scheme.",
"We consider the problem of encoding graphs with n vertices and m edges compactly supporting adjacency, neighborhood and degree queries in constant time in the @Q(logn)-bit word RAM model. The adjacency query asks whether there is an edge between two vertices, the neighborhood query reports the neighbors of a given vertex in constant time per neighbor, and the degree query reports the number of incident edges to a given vertex. We study the problem in the context of succinctness, where the goal is to achieve the optimal space requirement as a function of n and m, to within lower order terms. We prove a lower bound in the cell probe model indicating it is impossible to achieve the information-theory lower bound up to lower order terms unless the graph is either too sparse (namely, m=o(n^@d) for any constant @d>0) or too dense (namely m=@w(n^2^-^@d) for any constant @d>0). Furthermore, we present a succinct encoding of graphs supporting aforementioned queries in constant time. The space requirement of the encoding is within a multiplicative 1+@e factor of the information-theory lower bound for any arbitrarily small constant @e>0. This is the best achievable space bound according to our lower bound where it applies. The space requirement of the representation achieves the information-theory lower bound tightly within lower order terms where the graph is very sparse (m=o(n^@d) for any constant @d>0), or very dense (m>n^2 lg^1^-^@dn for an arbitrarily small constant @d>0).",
"We propose graph encryption schemes that efficiently support approximate shortest distance queries on large-scale encrypted graphs. Shortest distance queries are one of the most fundamental graph operations and have a wide range of applications. Using such graph encryption schemes, a client can outsource large-scale privacy-sensitive graphs to an untrusted server without losing the ability to query it. Other applications include encrypted graph databases and controlled disclosure systems. We propose GRECS (stands for GRaph EnCryption for approximate Shortest distance queries) which includes three oracle encryption schemes that are provably secure against any semi-honest server. Our first construction makes use of only symmetric-key operations, resulting in a computationally-efficient construction. Our second scheme makes use of somewhat-homomorphic encryption and is less computationally-efficient but achieves optimal communication complexity (i.e. uses a minimal amount of bandwidth). Finally, our third scheme is both computationally-efficient and achieves optimal communication complexity at the cost of a small amount of additional leakage. We implemented and evaluated the efficiency of our constructions experimentally. The experiments demonstrate that our schemes are efficient and can be applied to graphs that scale up to 1.6 million nodes and 11 million edges."
]
} |
1901.11267 | 2949631225 | Lifecycle models for research data are often abstract and simple. This comes at the danger of oversimplifying the complex concepts of research data management. The analysis of 90 different lifecycle models lead to two approaches to assess the quality of these models. While terminological issues make direct comparisons of models hard, an empirical evaluation seems possible. | @cite_8 shows an approach very similar to ours: Based on a survey of lifecycle models, an abstract data lifecycle model is derived and a classification scheme is developed. In contrast to @cite_8 , we do not define a lifecycle model but a common scheme shared by all found lifecycle models. One of features by which @cite_8 classifies, is the distinction between prescriptive and descriptive models, which comes very close to our proposal to classify along the purpose the model was designed for. Our method is more focused on evaluation and the resulting classification is therefore more fine-grained with regard to that. @cite_8 provides more classifications of features, of which some are irrelevant for evaluation (e.g. the distinction betwen homogeneous and heterogeneous lifecycles). | {
"cite_N": [
"@cite_8"
],
"mid": [
"2143774383"
],
"abstract": [
"This paper studies the problem of leveraging computationally intensive classification algorithms for large scale text categorization problems. We propose a hierarchical approach which decomposes the classification problem into a coarse level task and a fine level task. A simple yet scalable classifier is applied to perform the coarse level classification while a more sophisticated model is used to separate classes at the fine level. However, instead of relying on a human-defined hierarchy to decompose the problem, we we use a graph algorithm to discover automatically groups of highly similar classes. As an illustrative example, we apply our approach to real-world industrial data from eBay, a major e-commerce site where the goal is to classify live items into a large taxonomy of categories. In such industrial setting, classification is very challenging due to the number of classes, the amount of training data, the size of the feature space and the real-world requirements on the response time. We demonstrate through extensive experimental evaluation that (1) the proposed hierarchical approach is superior to flat models, and (2) the data-driven extraction of latent groups works significantly better than the existing human-defined hierarchy."
]
} |
1901.11267 | 2949631225 | Lifecycle models for research data are often abstract and simple. This comes at the danger of oversimplifying the complex concepts of research data management. The analysis of 90 different lifecycle models lead to two approaches to assess the quality of these models. While terminological issues make direct comparisons of models hard, an empirical evaluation seems possible. | @cite_10 , @cite_1 and @cite_7 are alike to @cite_8 in the approach to review existing models and deriving an own lifecycle model based on a gap analysis. None of the three publications offer generic and empirical evaluation criteria or a metamodel for the existing models. Their lifecycle model is designed to supersede the existing approaches for a specific context. | {
"cite_N": [
"@cite_1",
"@cite_10",
"@cite_7",
"@cite_8"
],
"mid": [
"2786674049",
"2165636119",
"2122778642",
"2484703597"
],
"abstract": [
"In this paper, we present the results of long-term research conducted in order to study the contribution made by software models based on the Unified Modeling Language (UML) to the comprehensibility of Java source-code deprived of comments. We have conducted 12 controlled experiments in different experimental contexts and on different sites with participants with different levels of expertise (i.e., Bachelor’s, Master’s, and PhD students and software practitioners from Italy and Spain). A total of 333 observations were obtained from these experiments. The UML models in our experiments were those produced in the analysis and design phases. The models produced in the analysis phase were created with the objective of abstracting the environment in which the software will work (i.e., the problem domain), while those produced in the design phase were created with the goal of abstracting implementation aspects of the software (i.e., the solution application domain). Source-code comprehensibility was assessed with regard to correctness of understanding, time taken to accomplish the comprehension tasks, and efficiency as regards accomplishing those tasks. In order to study the global effect of UML models on source-code comprehensibility, we aggregated results from the individual experiments using a meta-analysis. We made every effort to account for the heterogeneity of our experiments when aggregating the results obtained from them. The overall results suggest that the use of UML models affects the comprehensibility of source-code, when it is deprived of comments. Indeed, models produced in the analysis phase might reduce source-code comprehensibility, while increasing the time taken to complete comprehension tasks. That is, browsing source code and this kind of models together negatively impacts on the time taken to complete comprehension tasks without having a positive effect on the comprehensibility of source code. One plausible justification for this is that the UML models produced in the analysis phase focus on the problem domain. That is, models produced in the analysis phase say nothing about source code and there should be no expectation that they would, in any way, be beneficial to comprehensibility. On the other hand, UML models produced in the design phase improve source-code comprehensibility. One possible justification for this result is that models produced in the design phase are more focused on implementation details. Therefore, although the participants had more material to read and browse, this additional effort was paid back in the form of an improved comprehension of source code.",
"In this work, we address the problem of joint modeling of text and citations in the topic modeling framework. We present two different models called the Pairwise-Link-LDA and the Link-PLSA-LDA models. The Pairwise-Link-LDA model combines the ideas of LDA [4] and Mixed Membership Block Stochastic Models [1] and allows modeling arbitrary link structure. However, the model is computationally expensive, since it involves modeling the presence or absence of a citation (link) between every pair of documents. The second model solves this problem by assuming that the link structure is a bipartite graph. As the name indicates, Link-PLSA-LDA model combines the LDA and PLSA models into a single graphical model. Our experiments on a subset of Citeseer data show that both these models are able to predict unseen data better than the baseline model of Erosheva and Lafferty [8], by capturing the notion of topical similarity between the contents of the cited and citing documents. Our experiments on two different data sets on the link prediction task show that the Link-PLSA-LDA model performs the best on the citation prediction task, while also remaining highly scalable. In addition, we also present some interesting visualizations generated by each of the models.",
"When we write or prepare to write a research paper, we always have appropriate references in mind. However, there are most likely references we have missed and should have been read and cited. As such a good citation recommendation system would not only improve our paper but, overall, the efficiency and quality of literature search. Usually, a citation's context contains explicit words explaining the citation. Using this, we propose a method that \"translates\" research papers into references. By considering the citations and their contexts from existing papers as parallel data written in two different \"languages\", we adopt the translation model to create a relationship between these two \"vocabularies\". Experiments on both CiteSeer and CiteULike dataset show that our approach outperforms other baseline methods and increase the precision, recall and f-measure by at least 5 to 10 , respectively. In addition, our approach runs much faster in the both training and recommending stage, which proves the effectiveness and the scalability of our work.",
"Many organizations maintain textual process descriptions alongside graphical process models. The purpose is to make process information accessible to various stakeholders, including those who are not familiar with reading and interpreting the complex execution logic of process models. Despite this merit, there is a clear risk that model and text become misaligned when changes are not applied to both descriptions consistently. For organizations with hundreds of different processes, the effort required to identify and clear up such conflicts is considerable. To support organizations in keeping their process descriptions consistent, we present an approach to automatically identify inconsistencies between a process model and a corresponding textual description. Our approach detects cases where the two process representations describe activities in different orders and detect process model activities not contained in the textual description. A quantitative evaluation with 53 real-life model-text pairs demonstrates that our approach accurately identifies inconsistencies between model and text. HighlightsWe propose an approach to detect conflicts between textual and model-based process descriptions.The approach is fully automatic based on tailored natural language processing techniques.Quantitative evaluation demonstrates the applicability of the approach on real-life data."
]
} |
1901.11267 | 2949631225 | Lifecycle models for research data are often abstract and simple. This comes at the danger of oversimplifying the complex concepts of research data management. The analysis of 90 different lifecycle models lead to two approaches to assess the quality of these models. While terminological issues make direct comparisons of models hard, an empirical evaluation seems possible. | @cite_1 and @cite_7 both propose a lifecycle model for Big Data. Although they model the same phenomena, the models are not similar. While @cite_7 does not describe evaluation criteria of the model, @cite_1 proposes the 6Vs of Big Data (Value, Volume, Variety, Velocity, Variability, Veracity) as a base to evaluate data lifecycle models in the context of Big Data. This evaluation carried out in @cite_17 is also applied to evaluate other data lifecycle models for their aptness to describe Big Data challenges. This evaluation is the most rigorous we found in the literature, but it is limited to the context of Big Data and itself is based on a theoretical concept instead of empirical evaluation. | {
"cite_N": [
"@cite_1",
"@cite_7",
"@cite_17"
],
"mid": [
"2407478184",
"2558036245",
"2591999683"
],
"abstract": [
"As science becomes more data-intensive and collaborative, researchers increasingly use larger and more complex data to answer research questions. The capacity of storage infrastructure, the increased sophistication and deployment of sensors, the ubiquitous availability of computer clusters, the development of new analysis techniques, and larger collaborations allow researchers to address grand societ al challenges in a way that is unprecedented. In parallel, research data repositories have been built to host research data in response to the requirements of sponsors that research data be publicly available. Libraries are re-inventing themselves to respond to a growing demand to manage, store, curate and preserve the data produced in the course of publicly funded research. As librarians and data managers are developing the tools and knowledge they need to meet these new expectations, they inevitably encounter conversations around Big Data. This paper explores definitions of Big Data that have coalesced in the last decade around four commonly mentioned characteristics: volume, variety, velocity, and veracity. We highlight the issues associated with each characteristic, particularly their impact on data management and curation. We use the methodological framework of the data life cycle model, assessing two models developed in the context of Big Data projects and find them lacking. We propose a Big Data life cycle model that includes activities focused on Big Data and more closely integrates curation with the research life cycle. These activities include planning, acquiring, preparing, analyzing, preserving, and discovering, with describing the data and assuring quality being an integral part of each activity. We discuss the relationship between institutional data curation repositories and new long-term data resources associated with high performance computing centers, and reproducibility in computational science. We apply this model by mapping the four characteristics of Big Data outlined above to each of the activities in the model. This mapping produces a set of questions that practitioners should be asking in a Big Data project",
"A huge amount of data is constantly being produced in the world. Data coming from the IoT, from scientific simulations, or from any other field of the eScience, are accumulated over historical data sets and set up the seed for future Big Data processing, with the final goal to generate added value and discover knowledge. In such computing processes, data are the main resource, however, organizing and managing data during their entire life cycle becomes a complex research topic. As part of this, Data LifeCycle (DLC) models have been proposed to efficiently organize large and complex data sets, from creation to consumption, in any field, and any scale, for an effective data usage and big data exploitation. 2. Several DLC frameworks can be found in the literature, each one defined for specific environments and scenarios. However, we realized that there is no global and comprehensive DLC model to be easily adapted to different scientific areas. For this reason, in this paper we describe the Comprehensive Scenario Agnostic Data LifeCycle (COSA-DLC) model, a DLC model which: i) is proved to be comprehensive as it addresses the 6Vs challenges (namely Value, Volume, Variety, Velocity, Variability and Veracity, and ii), it can be easily adapted to any particular scenario and, therefore, fit the requirements of a specific scientific field. In this paper we also include two use cases to illustrate the ease of the adaptation in different scenarios. We conclude that the comprehensive scenario agnostic DLC model provides several advantages, such as facilitating global data management, organization and integration, easing the adaptation to any kind of scenario, guaranteeing good data quality levels and, therefore, saving design time and efforts for the scientific and industrial communities.",
"There is a vast amount of data being generated every day in the world, coming from a variety of sources, with different formats, quality levels, etc. This new data, together with the archived historical data, constitute the seed for future knowledge discovery and value generation in several fields of eScience. Discovering value from data is a complex computing process where data is the key resource, not only during its processing, but also during its entire life cycle. However, there is still a huge concern about how to organize and manage this data in all fields, and at all scales, for efficient usage and exploitation during all data life cycles. Although several specific Data LifeCycle (DLC) models have been recently defined for particular scenarios, we argue that there is no global and comprehensive DLC framework to be widely used in different fields. For this reason, in this paper we present and describe a comprehensive scenario agnostic Data LifeCycle (COSA-DLC) model successfully addressing all challenges included in the 6Vs, namely Value, Volume, Variety, Velocity, Variability and Veracity, not tailored to any specific environment, but easy to be adapted to fit the requirements of any particular field. We conclude that a comprehensive scenario agnostic DLC model provides several advantages, such as facilitating global data organization and integration, easing the adaptation to any kind of scenario, guaranteeing good quality data levels, and helping save design time and efforts for the research and industrial communities."
]
} |
1901.11267 | 2949631225 | Lifecycle models for research data are often abstract and simple. This comes at the danger of oversimplifying the complex concepts of research data management. The analysis of 90 different lifecycle models lead to two approaches to assess the quality of these models. While terminological issues make direct comparisons of models hard, an empirical evaluation seems possible. | @cite_20 provides a scoped review of 301 articles and 10 companion documents discussing research data management practices in academic institutions between 1995 and 2016. The review is not limited to, but includes publications discussing data lifecycle models. The discussion includes the observation, that of the papers reviewed, only a view provided empirical evidence for their results, which is in accordance to our findings. The study classifies the papers based on the UK data lifecycle https: www.ukdataservice.ac.uk manage-data lifecycle , which fortunately is preserved as an attachement to this paper (its "official" version has changed since the original publication). | {
"cite_N": [
"@cite_20"
],
"mid": [
"2617597628"
],
"abstract": [
"Objective The purpose of this study is to describe the volume, topics, and methodological nature of the existing research literature on research data management in academic institutions. Materials and methods We conducted a scoping review by searching forty literature databases encompassing a broad range of disciplines from inception to April 2016. We included all study types and data extracted on study design, discipline, data collection tools, and phase of the research data lifecycle. Results We included 301 articles plus 10 companion reports after screening 13,002 titles and abstracts and 654 full-text articles. Most articles (85 ) were published from 2010 onwards and conducted within the sciences (86 ). More than three-quarters of the articles (78 ) reported methods that included interviews, cross-sectional, or case studies. Most articles (68 ) included the Giving Access to Data phase of the UK Data Archive Research Data Lifecycle that examines activities such as sharing data. When studies were grouped into five dominant groupings (Stakeholder, Data, Library, Tool Device, and Publication), data quality emerged as an integral element. Conclusion Most studies relied on self-reports (interviews, surveys) or accounts from an observer (case studies) and we found few studies that collected empirical evidence on activities amongst data producers, particularly those examining the impact of research data management interventions. As well, fewer studies examined research data management at the early phases of research projects. The quality of all research outputs needs attention, from the application of best practices in research data management studies, to data producers depositing data in repositories for long-term use."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | In the absence of external parallel data, one polylingual DSM which has recently proved worthy (and that we use as a baseline in our experiments) is (LRI -- @cite_2 ), the polylingual extension of the (RI) method @cite_22 . RI is a context-counting model belonging to the family of random projection methods, and is considered a cheaper approximation of LSA @cite_47 . LRI is designed so that the orthogonality of the projection base is maximized, which allows to preserve sparsity and maximize the contribution of the information conveyed by the features shared across languages. | {
"cite_N": [
"@cite_47",
"@cite_22",
"@cite_2"
],
"mid": [
"1983623418",
"2963611534",
"1504212872"
],
"abstract": [
"A standard approach to cross-language information retrieval (CLIR) uses Latent Semantic Analysis (LSA) in conjunction with a multilingual parallel aligned corpus. This approach has been shown to be successful in identifying similar documents across languages - or more precisely, retrieving the most similar document in one language to a query in another language. However, the approach has severe drawbacks when applied to a related task, that of clustering documents \"language-independently\", so that documents about similar topics end up closest to one another in the semantic space regardless of their language. The problem is that documents are generally more similar to other documents in the same language than they are to documents in a different language, but on the same topic. As a result, when using multilingual LSA, documents will in practice cluster by language, not by topic. We propose a novel application of PARAFAC2 (which is a variant of PARAFAC, a multi-way generalization of the singular value decomposition [SVD]) to overcome this problem. Instead of forming a single multilingual term-by-document matrix which, under LSA, is subjected to SVD, we form an irregular three-way array, each slice of which is a separate term-by-document matrix for a single language in the parallel corpus. The goal is to compute an SVD for each language such that V (the matrix of right singular vectors) is the same across all languages. Effectively, PARAFAC2 imposes the constraint, not present in standard LSA, that the \"concepts\" in all documents in the parallel corpus are the same regardless of language. Intuitively, this constraint makes sense, since the whole purpose of using a parallel corpus is that exactly the same concepts are expressed in the translations. We tested this approach by comparing the performance of PARAFAC2 with standard LSA in solving a particular CLIR problem. From our results, we conclude that PARAFAC2 offers a very promising alternative to LSA not only for multilingual document clustering, but also for solving other problems in cross-language information retrieval.",
"Despite the availability of a huge amount of video data accompanied by descriptive texts, it is not always easy to exploit the information contained in natural language in order to automatically recognize video concepts. Towards this goal, in this paper we use textual cues as means of supervision, introducing two weakly supervised techniques that extend the Multiple Instance Learning (MIL) framework: the Fuzzy Sets Multiple Instance Learning (FSMIL) and the Probabilistic Labels Multiple Instance Learning (PLMIL). The former encodes the spatio-temporal imprecision of the linguistic descriptions with Fuzzy Sets, while the latter models different interpretations of each description's semantics with Probabilistic Labels, both formulated through a convex optimization algorithm. In addition, we provide a novel technique to extract weak labels in the presence of complex semantics, that consists of semantic similarity computations. We evaluate our methods on two distinct problems, namely face and action recognition, in the challenging and realistic setting of movies accompanied by their screenplays, contained in the COGNIMUSE database. We show that, on both tasks, our method considerably outperforms a state-of-the-art weakly supervised approach, as well as other baselines.",
"Objective : Development of a general natural-language processor that identifies clinical information in narrative reports and maps that information into a structured representation containing clinical terms. @PARASPLIT Design : The natural-language processor provides three phases of processing, all of which are driven by different knowledge sources. The first phase performs the parsing. It identifies the structure of the text through use of a grammar that defines semantic patterns and a target form. The second phase, regularization, standardizes the terms in the initial target structure via a compositional mapping of multi-word phrases. The third phase, encoding, maps the terms to a controlled vocabulary. Radiology is the test domain for the processor and the target structure is a formal model for representing clinical information in that domain. @PARASPLIT Measurements : The impression sections of 230 radiology reports were encoded by the processor. Results of an automated query of the resultant database for the occurrences of four diseases were compared with the analysis of a panel of three physicians to determine recall and precision. @PARASPLIT Results : Without training specific to the four diseases, recall and precision of the system(combined effect of the processor and query generator) were 70 and 87 . Training of the query component increased recall to 85 without changing precision."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | Another method that requires external multilingual resources (specifically: a word translation oracle) is (CL-SCL -- @cite_39 ). CL-SCL relies on solving auxiliary prediction problems, which consist in discovering hidden correlations between terms in a language. This is achieved by binary classifiers trained to predict the presence of highly discriminative terms ( pivots'') given the other terms in the document. The cross-lingual aspect is addressed by imposing that pivot terms are aligned (i.e., translations of each other) across languages, which requires a word translation oracle. A stronger, more recent variant of CL-SCL (which we also compare against in our experiments) is (DCI -- @cite_44 ). DCI derives term representations in a vector space common to all languages where each dimension reflects its distributional correspondence (as quantified by a distributional correspondence function'') to a pivot. Machine Translation (MT) represents an appealing tool to solve PLC, and several PLC methods are indeed based on the use of MT services @cite_40 @cite_42 . However, the drawback of these methods is reduced generality, since it is not always the case that quality MT tools are both (i) available for the required language combinations, and (ii) free to use. | {
"cite_N": [
"@cite_44",
"@cite_40",
"@cite_42",
"@cite_39"
],
"mid": [
"2296116776",
"2027570514",
"2573834658",
"2953109491"
],
"abstract": [
"This paper describes a technique to exploit multiple pivot languages when using machine translation (MT) on language pairs with scarce bilingual resources, or where no translation system for a language pair is available. The principal idea is to generate intermediate translations in several pivot languages, translate them separately into the target language, and generate a consensus translation out of these using MT system combination techniques. Our technique can also be applied when a translation system for a language pair is available, but is limited in its translation accuracy because of scarce resources. Using statistical MT systems for the 11 different languages of Europarl, we show experimentally that a direct translation system can be replaced by this pivot approach without a loss in translation quality if about six pivot languages are available. Furthermore, we can already improve an existing MT system by adding two pivot systems to it. The maximum improvement was found to be 1.4 abs. in BLEU in our experiments for 8 or more pivot languages.",
"Cross-lingual adaptation is a special case of domain adaptation and refers to the transfer of classification knowledge between two languages. In this article we describe an extension of Structural Correspondence Learning (SCL), a recently proposed algorithm for domain adaptation, for cross-lingual adaptation in the context of text classification. The proposed method uses unlabeled documents from both languages, along with a word translation oracle, to induce a cross-lingual representation that enables the transfer of classification knowledge from the source to the target language. The main advantages of this method over existing methods are resource efficiency and task specificity. We conduct experiments in the area of cross-language topic and sentiment classification involving English as source language and German, French, and Japanese as target languages. The results show a significant improvement of the proposed method over a machine translation baseline, reducing the relative error due to cross-lingual adaptation by an average of 30p (topic classification) and 59p (sentiment classification). We further report on empirical analyses that reveal insights into the use of unlabeled data, the sensitivity with respect to important hyperparameters, and the nature of the induced cross-lingual word correspondences.",
"We propose an approach to build a neural machine translation system with no supervised resources (i.e., no parallel corpora) using multimodal embedded representation over texts and images. Based on the assumption that text documents are often likely to be described with other multimedia information (e.g., images) somewhat related to the content, we try to indirectly estimate the relevance between two languages. Using multimedia as the \"pivot\", we project all modalities into one common hidden space where samples belonging to similar semantic concepts should come close to each other, whatever the observed space of each sample is. This modality-agnostic representation is the key to bridging the gap between different modalities. Putting a decoder on top of it, our network can flexibly draw the outputs from any input modality. Notably, in the testing phase, we need only source language texts as the input for translation. In experiments, we tested our method on two benchmarks to show that it can achieve reasonable translation performance. We compared and investigated several possible implementations and found that an end-to-end model that simultaneously optimized both rank loss in multimodal encoders and cross-entropy loss in decoders performed the best.",
"State-of-the-art natural language processing systems rely on supervision in the form of annotated data to learn competent models. These models are generally trained on data in a single language (usually English), and cannot be directly used beyond that language. Since collecting data in every language is not realistic, there has been a growing interest in cross-lingual language understanding (XLU) and low-resource cross-language transfer. In this work, we construct an evaluation set for XLU by extending the development and test sets of the Multi-Genre Natural Language Inference Corpus (MultiNLI) to 15 languages, including low-resource languages such as Swahili and Urdu. We hope that our dataset, dubbed XNLI, will catalyze research in cross-lingual sentence understanding by providing an informative standard evaluation task. In addition, we provide several baselines for multilingual sentence understanding, including two based on machine translation systems, and two that use parallel data to train aligned multilingual bag-of-words and LSTM encoders. We find that XNLI represents a practical and challenging evaluation suite, and that directly translating the test data yields the best performance among available baselines."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | Approaches to PLC based on deep learning focus on defining representations based on word embeddings which capture the semantic regularities in language while at the same time being aligned across languages. In order to produce aligned representations, though, deep learning approaches typically require the availability of external parallel corpora @cite_37 @cite_46 , bi-lingual lexicons @cite_34 , or machine translation tools @cite_3 . Recently, proposed a method to align monolingual word embedding spaces (as those produced by, e.g., Word2Vec @cite_48 ) from different languages without requiring parallel data. To this aim, @cite_7 proposed an adversarial training process in which a (in charge of mapping the source embeddings onto the target space) is trained to fool a from distinguishing the provenance of the embeddings, i.e., from understanding whether the embeddings it receives as input come from the (transformed) source or from the target space. After that, the mapping is refined by means of unsupervised techniques. Despite operating without parallel resources, @cite_7 obtained state-of-the-art multilingual mappings, which they later made publicly available https: github.com facebookresearch MUSE and which we use as a further baseline in our experiments of Section . | {
"cite_N": [
"@cite_37",
"@cite_7",
"@cite_48",
"@cite_3",
"@cite_46",
"@cite_34"
],
"mid": [
"2950682695",
"2762484717",
"2766182427",
"2229725139",
"2962824887",
"2767206889"
],
"abstract": [
"Cross-language learning allows us to use training data from one language to build models for a different language. Many approaches to bilingual learning require that we have word-level alignment of sentences from parallel corpora. In this work we explore the use of autoencoder-based methods for cross-language learning of vectorial word representations that are aligned between two languages, while not relying on word-level alignments. We show that by simply learning to reconstruct the bag-of-words representations of aligned sentences, within and between languages, we can in fact learn high-quality representations and do without word alignments. Since training autoencoders on word observations presents certain computational issues, we propose and compare different variations adapted to this setting. We also propose an explicit correlation maximizing regularizer that leads to significant improvement in the performance. We empirically investigate the success of our approach on the problem of cross-language test classification, where a classifier trained on a given language (e.g., English) must learn to generalize to a different language (e.g., German). These experiments demonstrate that our approaches are competitive with the state-of-the-art, achieving up to 10-14 percentage point improvements over the best reported results on this task.",
"State-of-the-art methods for learning cross-lingual word embeddings have relied on bilingual dictionaries or parallel corpora. Recent studies showed that the need for parallel data supervision can be alleviated with character-level information. While these methods showed encouraging results, they are not on par with their supervised counterparts and are limited to pairs of languages sharing a common alphabet. In this work, we show that we can build a bilingual dictionary between two languages without using any parallel corpora, by aligning monolingual word embedding spaces in an unsupervised way. Without using any character information, our model even outperforms existing supervised methods on cross-lingual tasks for some language pairs. Our experiments demonstrate that our method works very well also for distant language pairs, like English-Russian or English-Chinese. We finally describe experiments on the English-Esperanto low-resource language pair, on which there only exists a limited amount of parallel data, to show the potential impact of our method in fully unsupervised machine translation. Our code, embeddings and dictionaries are publicly available.",
"In spite of the recent success of neural machine translation (NMT) in standard benchmarks, the lack of large parallel corpora poses a major practical problem for many language pairs. There have been several proposals to alleviate this issue with, for instance, triangulation and semi-supervised learning techniques, but they still require a strong cross-lingual signal. In this work, we completely remove the need of parallel data and propose a novel method to train an NMT system in a completely unsupervised manner, relying on nothing but monolingual corpora. Our model builds upon the recent work on unsupervised embedding mappings, and consists of a slightly modified attentional encoder-decoder model that can be trained on monolingual corpora alone using a combination of denoising and backtranslation. Despite the simplicity of the approach, our system obtains 15.56 and 10.21 BLEU points in WMT 2014 French-to-English and German-to-English translation. The model can also profit from small parallel corpora, and attains 21.81 and 15.24 points when combined with 100,000 parallel sentences, respectively. Our implementation is released as an open source project.",
"We propose a new model for learning bilingual word representations from nonparallel document-aligned data. Following the recent advances in word representation learning, our model learns dense real-valued word vectors, that is, bilingual word embeddings (BWEs). Unlike prior work on inducing BWEs which heavily relied on parallel sentence-aligned corpora and or readily available translation resources such as dictionaries, the article reveals that BWEs may be learned solely on the basis of document-aligned comparable data without any additional lexical resources nor syntactic information. We present a comparison of our approach with previous state-of-the-art models for learning bilingual word representations from comparable data that rely on the framework of multilingual probabilistic topic modeling (MuPTM), as well as with distributional local context-counting models. We demonstrate the utility of the induced BWEs in two semantic tasks: (1) bilingual lexicon extraction, (2) suggesting word translations in context for polysemous words. Our simple yet effective BWE-based models significantly outperform the MuPTM-based and context-counting representation models from comparable data as well as prior BWE-based models, and acquire the best reported results on both tasks for all three tested language pairs.",
"In spite of the recent success of neural machine translation (NMT) in standard benchmarks, the lack of large parallel corpora poses a major practical problem for many language pairs. There have been several proposals to alleviate this issue with, for instance, triangulation and semi-supervised learning techniques, but they still require a strong cross-lingual signal. In this work, we completely remove the need of parallel data and propose a novel method to train an NMT system in a completely unsupervised manner, relying on nothing but monolingual corpora. Our model builds upon the recent work on unsupervised embedding mappings, and consists of a slightly modified attentional encoder-decoder model that can be trained on monolingual corpora alone using a combination of denoising and backtranslation. Despite the simplicity of the approach, our system obtains 15.56 and 10.21 BLEU points in WMT 2014 French-to-English and German-to-English translation. The model can also profit from small parallel corpora, and attains 21.81 and 15.24 points when combined with 100,000 parallel sentences, respectively.",
"Existing approaches to neural machine translation condition each output word on previously generated outputs. We introduce a model that avoids this autoregressive property and produces its outputs in parallel, allowing an order of magnitude lower latency during inference. Through knowledge distillation, the use of input token fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer network used as a teacher. We demonstrate substantial cumulative improvements associated with each of the three aspects of our training strategy, and validate our approach on IWSLT 2016 English-German and two WMT language pairs. By sampling fertilities in parallel at inference time, our non-autoregressive model achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English-Romanian."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | Funnelling is reminiscent of the (a.k.a. stacking'') method for ensemble learning @cite_36 . Let us discuss their commonalities and differences. | {
"cite_N": [
"@cite_36"
],
"mid": [
"1953606363"
],
"abstract": [
"This paper proposes an ensemble method for multilabel classification. The RAndom k-labELsets (RAKEL) algorithm constructs each member of the ensemble by considering a small random subset of labels and learning a single-label classifier for the prediction of each element in the powerset of this subset. In this way, the proposed algorithm aims to take into account label correlations using single-label classifiers that are applied on subtasks with manageable number of labels and adequate number of examples per label. Experimental results on common multilabel domains involving protein, document and scene classification show that better performance can be achieved compared to popular multilabel classification approaches."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | Common to stacking and funnelling is the presence of an ensemble of @math base classifiers, typically trained on traditional'' vectorial representations, and the presence of a single meta-classifier that operates on vectors of base-classifier outputs. Common to stacking and is also the use of @math -fold cross-validation in order to generate the vectors of base-classifier outputs that are used to train the meta-classifier. (Variants of stacking in which @math -fold cross-validation is not used, and thus akin to , also exist @cite_26 .) | {
"cite_N": [
"@cite_26"
],
"mid": [
"2023294425"
],
"abstract": [
"We empirically evaluate several state-of-the-art methods for constructing ensembles of heterogeneous classifiers with stacking and show that they perform (at best) comparably to selecting the best classifier from the ensemble by cross validation. Among state-of-the-art stacking methods, stacking with probability distributions and multi-response linear regression performs best. We propose two extensions of this method, one using an extended set of meta-level features and the other using multi-response model trees to learn at the meta-level. We show that the latter extension performs better than existing stacking approaches and better than selecting the best classifier by cross validation."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | However, a key difference between the two methods is that stacking (like other ensemble methods such as bagging @cite_10 and boosting @cite_35 ) deals with ( homogeneous'') scenarios in which all training documents can in principle be represented in the same feature space and can thus concur to training the same classifier; in turn, this classifier can be used for classifying all the unlabelled documents. In stacking, the base classifiers sometimes differ in terms of the learning algorithm used to train them @cite_26 @cite_41 , or in terms of the subsets of the training set which are used for training them @cite_14 . In other words, in these scenarios setting up an ensemble is a choice, and not a necessity. It is instead a necessity in the ( heterogeneous'') scenarios which funnelling deals with, where labelled documents of different types (in our case: languages) could otherwise concur in training the same classifier (since they lie in different feature spaces), and where unlabelled documents could not (for analogous reasons) be classified by the same classifier. | {
"cite_N": [
"@cite_35",
"@cite_14",
"@cite_26",
"@cite_41",
"@cite_10"
],
"mid": [
"1605688901",
"2099031744",
"1529840045",
"2168118654",
"2100805904"
],
"abstract": [
"Bagging and boosting are methods that generate a diverse ensemble of classifiers by manipulating the training data given to a “base” learning algorithm. Breiman has pointed out that they rely for their effectiveness on the instability of the base learning algorithm. An alternative approach to generating an ensemble is to randomize the internal decisions made by the base algorithm. This general approach has been studied previously by Ali and Pazzani and by Dietterich and Kong. This paper compares the effectiveness of randomization, bagging, and boosting for improving the performance of the decision-tree algorithm C4.5. The experiments show that in situations with little or no classification noise, randomization is competitive with (and perhaps slightly superior to) bagging but not as accurate as boosting. In situations with substantial classification noise, bagging is much better than boosting, and sometimes better than randomization.",
"Due to the globalization on the Web, many companies and institutions need to efficiently organize and search repositories containing multilingual documents. The management of these heterogeneous text collections increases the costs significantly because experts of different languages are required to organize these collections. Cross-language text categorization can provide techniques to extend existing automatic classification systems in one language to new languages without requiring additional intervention of human experts. In this paper, we propose a learning algorithm based on the EM scheme which can be used to train text classifiers in a multilingual environment. In particular, in the proposed approach, we assume that a predefined category set and a collection of labeled training data is available for a given language L sub 1 . A classifier for a different language L sub 2 is trained by translating the available labeled training set for L sub 1 to L sub 2 and by using an additional set of unlabeled documents from L sub 2 . This technique allows us to extract correct statistical properties of the language L sub 2 which are not completely available in automatically translated examples, because of the different characteristics of language L sub 1 and of the approximation of the translation process. Our experimental results show that the performance of the proposed method is very promising when applied on a test document set extracted from newsgroups in English and Italian.",
"Bagging and boosting are two of the most well-known ensemble learning methods due to their theoretical performance guarantees and strong experimental results. However, these algorithms have been used mainly in batch mode, i.e., they require the entire training set to be available at once and, in some cases, require random access to the data. In this paper, we present online versions of bagging and boosting that require only one pass through the training data. We build on previously presented work by describing some theoretical results. We also compare the online and batch algorithms experimentally in terms of accuracy and running time.",
"One of the potential advantages of multiple classifier systems is an increased robustness to noise and other imperfections in data. Previous experiments on classification noise have shown that bagging is fairly robust but that boosting is quite sensitive. Decorate is a recently introduced ensemble method that constructs diverse committees using artificial data. It has been shown to generally outperform both boosting and bagging when training data is limited. This paper compares the sensitivity of bagging, boosting, and Decorate to three types of imperfect data: missing features, classification noise, and feature noise. For missing data, Decorate is the most robust. For classification noise, bagging and Decorate are both robust, with bagging being slightly better than Decorate, while boosting is quite sensitive. For feature noise, all of the ensemble methods increase the resilience of the base classifier.",
"An ensemble consists of a set of individually trained classifiers (such as neural networks or decision trees) whose predictions are combined when classifying novel instances. Previous research has shown that an ensemble is often more accurate than any of the single classifiers in the ensemble. Bagging (Breiman, 1996c) and Boosting (Freund & Schapire, 1996; Schapire, 1990) are two relatively new but popular methods for producing ensembles. In this paper we evaluate these methods on 23 data sets using both neural networks and decision trees as our classification algorithm. Our results clearly indicate a number of conclusions. First, while Bagging is almost always more accurate than a single classifier, it is sometimes much less accurate than Boosting. On the other hand, Boosting can create ensembles that are less accurate than a single classifier - especially when using neural networks. Analysis indicates that the performance of the Boosting methods is dependent on the characteristics of the data set being examined. In fact, further results show that Boosting ensembles may overfit noisy data sets, thus decreasing its performance. Finally, consistent with previous studies, our work suggests that most of the gain in an ensemble's performance comes in the first few classifiers combined; however, relatively large gains can be seen up to 25 classifiers when Boosting decision trees."
]
} |
1901.11459 | 2963967134 | Cross-lingual Text Classification (CLC) consists of automatically classifying, according to a common set C of classes, documents each written in one of a set of languages L, and doing so more accurately than when “naively” classifying each document via its corresponding language-specific classifier. To obtain an increase in the classification accuracy for a given language, the system thus needs to also leverage the training examples written in the other languages. We tackle “multilabel” CLC via funnelling, a new ensemble learning method that we propose here. Funnelling consists of generating a two-tier classification system where all documents, irrespective of language, are classified by the same (second-tier) classifier. For this classifier, all documents are represented in a common, language-independent feature space consisting of the posterior probabilities generated by first-tier, language-dependent classifiers. This allows the classification of all test documents, of any language, to benefit from the information present in all training documents, of any language. We present substantial experiments, run on publicly available multilingual text collections, in which funnelling is shown to significantly outperform a number of state-of-the-art baselines. All code and datasets (in vector form) are made publicly available. | It certainly that exist in multilabel settings @cite_28 @cite_5 @cite_18 , which is not possible when (as customarily done) a multilabel classification task is solved as @math independent binary classification problems. In fact, for an unlabelled document @math the meta-classifier receives @math inputs from the base classifier which has classified @math , and returns @math outputs, which means that the input for class @math has a potential impact on the output for class @math , for every choice of @math and @math . For instance, the fact that for @math the posterior probability for class Skiing is high might bring additional evidence that @math belongs to class Snowboarding ; this could be the result of several training documents labelled by Snowboarding having, in their @math vectors, a high value for class Skiing . | {
"cite_N": [
"@cite_28",
"@cite_5",
"@cite_18"
],
"mid": [
"2963185791",
"2166886337",
"2560045542"
],
"abstract": [
"Multiclass classification problems such as image annotation can involve a large number of classes. In this context, confusion between classes can occur, and single label classification may be misleading. We provide in the present paper a general device that, given an unlabeled dataset and a score function defined as the minimizer of some empirical and convex risk, outputs a set of class labels, instead of a single one. Interestingly, this procedure does not require that the unlabeled dataset explores the whole classes. Even more, the method is calibrated to control the expected size of the output set while minimizing the classification risk. We show the statistical optimality of the procedure and establish rates of convergence under the Tsybakov margin condition. It turns out that these rates are linear on the number of labels. We apply our methodology to convex aggregation of confidence sets based on the V-fold cross validation principle also known as the superlearning principle. We illustrate the numerical performance of the procedure on real data and demonstrate in particular that with moderate expected size, w.r.t. the number of labels, the procedure provides significant improvement of the classification risk.",
"We propose a new problem formulation which is similar to, but more informative than, the binary multiple-instance learning problem. In this setting, we are given groups of instances (described by feature vectors) along with estimates of the fraction of positively-labeled instances per group. The task is to learn an instance level classifier from this information. That is, we are trying to estimate the unknown binary labels of individuals from knowledge of group statistics. We propose a principled probabilistic model to solve this problem that accounts for uncertainty in the parameters and in the unknown individual labels. This model is trained with an efficient MCMC algorithm. Its performance is demonstrated on both synthetic and real-world data arising in general object recognition.",
"Learning a classifier from groups of unlabeled data, only knowing, for each group, the proportions of data with particular labels, is an important branch of classification tasks that are conceivable in many practical applications. In this paper, we proposed a novel solution for the problem of learning with label proportions (LLP) based on nonparallel support vector machines, termed as proportion-NPSVM, which can improve the classifiers to be a pair of nonparallel classification hyperplanes. The unique property of our method is that it only needs to solve a pair of smaller quadratic programming problems. Moreover, it can efficiently incorporate the known group label proportions with the latent unknown observation labels into one optimization model under a large-margin framework. Compared to the existing approaches, there are several advantages shown as follows: 1) it does not need to make restrictive assumptions on the training data; 2) nonparallel classifiers can be achieved without computing the large inverse matrices; 3) the optimization model can be effectively solved by using the alternative strategy with SMO technique or SOR method; 4) proportion-NPSVM has better generalization ability. Sufficient experimental results on both binary-classes and multi-classes data sets show the efficiency of our proposed method in classification accuracy, which prove the state-of-the-art method for LLP problems compared with competing algorithms."
]
} |
1901.11382 | 2950001871 | In the big data era, the impetus to digitize the vast reservoirs of data trapped in unstructured scanned documents such as invoices, bank documents and courier receipts has gained fresh momentum. The scanning process often results in the introduction of artifacts such as background noise, blur due to camera motion, watermarkings, coffee stains, or faded text. These artifacts pose many readability challenges to current text recognition algorithms and significantly degrade their performance. Existing learning based denoising techniques require a dataset comprising of noisy documents paired with cleaned versions. In such scenarios, a model can be trained to generate clean documents from noisy versions. However, very often in the real world such a paired dataset is not available, and all we have for training our denoising model are unpaired sets of noisy and clean images. This paper explores the use of GANs to generate denoised versions of the noisy documents. In particular, where paired information is available, we formulate the problem as an image-to-image translation task i.e, translating a document from noisy domain ( i.e., background noise, blurred, faded, watermarked ) to a target clean document using Generative Adversarial Networks (GAN). However, in the absence of paired images for training, we employed CycleGAN which is known to learn a mapping between the distributions of the noisy images to the denoised images using unpaired data to achieve image-to-image translation for cleaning the noisy documents. We compare the performance of CycleGAN for document cleaning tasks using unpaired images with a Conditional GAN trained on paired data from the same dataset. Experiments were performed on a public document dataset on which different types of noise were artificially induced, results demonstrate that CycleGAN learns a more robust mapping from the space of noisy to clean documents. | Generative adversarial Network (GAN) @cite_6 is the idea that has taken deep learning by storm. It employs adversarial training which essentially means pitting two neural networks against each other. One is a generator while the other is a discriminator, where the former aims at producing data that are indistinguishable from real data while the latter tries to distinguish between real and fake data. The process eventually yields a generator with the ability to do a plethora of tasks efficiently such as image-to-image generation. Other notable applications where GANs have established their supermacy are representation learning, image editing, art generation, music generation etc. @cite_20 @cite_7 @cite_21 @cite_3 @cite_17 . | {
"cite_N": [
"@cite_7",
"@cite_21",
"@cite_3",
"@cite_6",
"@cite_20",
"@cite_17"
],
"mid": [
"2737057113",
"2808763756",
"2753491454",
"2963474063",
"2616969219",
"2963376432"
],
"abstract": [
"Generative Adversarial Networks (GANs) have been shown to be able to sample impressively realistic images. GAN training consists of a saddle point optimization problem that can be thought of as an adversarial game between a generator which produces the images, and a discriminator, which judges if the images are real. Both the generator and the discriminator are commonly parametrized as deep convolutional neural networks. The goal of this paper is to disentangle the contribution of the optimization procedure and the network parametrization to the success of GANs. To this end we introduce and study Generative Latent Optimization (GLO), a framework to train a generator without the need to learn a discriminator, thus avoiding challenging adversarial optimization problems. We show experimentally that GLO enjoys many of the desirable properties of GANs: learning from large data, synthesizing visually-appealing samples, interpolating meaningfully between samples, and performing linear arithmetic with noise vectors.",
"Generative Adversarial Network (GAN) is a prominent generative model that are widely used in various applications. Recent studies have indicated that it is possible to obtain fake face images with a high visual quality based on this novel model. If those fake faces are abused in image tampering, it would cause some potential moral, ethical and legal problems. In this paper, therefore, we first propose a Convolutional Neural Network (CNN) based method to identify fake face images generated by the current best method [20], and provide experimental evidences to show that the proposed method can achieve satisfactory results with an average accuracy over 99.4 . In addition, we provide comparative results evaluated on some variants of the proposed CNN architecture, including the high pass filter, the number of the layer groups and the activation function, to further verify the rationality of our method.",
"A Triangle Generative Adversarial Network ( @math -GAN) is developed for semi-supervised cross-domain joint distribution matching, where the training data consists of samples from each domain, and supervision of domain correspondence is provided by only a few paired samples. @math -GAN consists of four neural networks, two generators and two discriminators. The generators are designed to learn the two-way conditional distributions between the two domains, while the discriminators implicitly define a ternary discriminative function, which is trained to distinguish real data pairs and two kinds of fake data pairs. The generators and discriminators are trained together using adversarial learning. Under mild assumptions, in theory the joint distributions characterized by the two generators concentrate to the data distribution. In experiments, three different kinds of domain pairs are considered, image-label, image-image and image-attribute pairs. Experiments on semi-supervised image classification, image-to-image translation and attribute-based image generation demonstrate the superiority of the proposed approach.",
"A Triangle Generative Adversarial Network ( @math -GAN) is developed for semi-supervised cross-domain joint distribution matching, where the training data consists of samples from each domain, and supervision of domain correspondence is provided by only a few paired samples. @math -GAN consists of four neural networks, two generators and two discriminators. The generators are designed to learn the two-way conditional distributions between the two domains, while the discriminators implicitly define a ternary discriminative function, which is trained to distinguish real data pairs and two kinds of fake data pairs. The generators and discriminators are trained together using adversarial learning. Under mild assumptions, in theory the joint distributions characterized by the two generators concentrate to the data distribution. In experiments, three different kinds of domain pairs are considered, image-label, image-image and image-attribute pairs. Experiments on semi-supervised image classification, image-to-image translation and attribute-based image generation demonstrate the superiority of the proposed approach.",
"Generative adversarial networks (GANs) have great successes on synthesizing data. However, the existing GANs restrict the discriminator to be a binary classifier, and thus limit their learning capacity for tasks that need to synthesize output with rich structures such as natural language descriptions. In this paper, we propose a novel generative adversarial network, RankGAN, for generating high-quality language descriptions. Rather than training the discriminator to learn and assign absolute binary predicate for individual data sample, the proposed RankGAN is able to analyze and rank a collection of human-written and machine-written sentences by giving a reference group. By viewing a set of data samples collectively and evaluating their quality through relative ranking scores, the discriminator is able to make better assessment which in turn helps to learn a better generator. The proposed RankGAN is optimized through the policy gradient technique. Experimental results on multiple public datasets clearly demonstrate the effectiveness of the proposed approach.",
"Generative Adversarial Networks (GANs) have achieved remarkable results in the task of generating realistic natural images. In most applications, GAN models share two aspects in common. On the one hand, GANs training involves solving a challenging saddle point optimization problem, interpreted as an adversarial game between a generator and a discriminator functions. On the other hand, the generator and the discriminator are parametrized in terms of deep convolutional neural networks. The goal of this paper is to disentangle the contribution of these two factors to the success of GANs. In particular, we introduce Generative Latent Optimization (GLO), a framework to train deep convolutional generators without using discriminators, thus avoiding the instability of adversarial optimization problems. Throughout a variety of experiments, we show that GLO enjoys many of the desirable properties of GANs: learning from large data, synthesizing visually-appealing samples, interpolating meaningfully between samples, and performing linear arithmetic with noise vectors."
]
} |
1901.11382 | 2950001871 | In the big data era, the impetus to digitize the vast reservoirs of data trapped in unstructured scanned documents such as invoices, bank documents and courier receipts has gained fresh momentum. The scanning process often results in the introduction of artifacts such as background noise, blur due to camera motion, watermarkings, coffee stains, or faded text. These artifacts pose many readability challenges to current text recognition algorithms and significantly degrade their performance. Existing learning based denoising techniques require a dataset comprising of noisy documents paired with cleaned versions. In such scenarios, a model can be trained to generate clean documents from noisy versions. However, very often in the real world such a paired dataset is not available, and all we have for training our denoising model are unpaired sets of noisy and clean images. This paper explores the use of GANs to generate denoised versions of the noisy documents. In particular, where paired information is available, we formulate the problem as an image-to-image translation task i.e, translating a document from noisy domain ( i.e., background noise, blurred, faded, watermarked ) to a target clean document using Generative Adversarial Networks (GAN). However, in the absence of paired images for training, we employed CycleGAN which is known to learn a mapping between the distributions of the noisy images to the denoised images using unpaired data to achieve image-to-image translation for cleaning the noisy documents. We compare the performance of CycleGAN for document cleaning tasks using unpaired images with a Conditional GAN trained on paired data from the same dataset. Experiments were performed on a public document dataset on which different types of noise were artificially induced, results demonstrate that CycleGAN learns a more robust mapping from the space of noisy to clean documents. | Image-to-image translation is the task of mapping images in source domain to images in target domain such as converting sketches into photographs, grayscale images to color images etc. The aim is to generate the target distribution given the source distribution. Prior work in the field of GANs such as Conditional GAN @cite_4 forces the image produced by generator to be conditioned on the output which allows for optimal translations. However, earlier GANs require one-to-one mapping of images between source and target domain i.e., a paired dataset. In case of documents, it is not possible to always have cleaned documents corresponding to each noisy document. This persuaded us to explore unpaired image-to-image translation methods, e.g. Dual-GAN @cite_19 which uses dual learning and CycleGAN @cite_1 which makes use of cyclic-consistency loss to achieve unpaired image-to-image translation. | {
"cite_N": [
"@cite_1",
"@cite_19",
"@cite_4"
],
"mid": [
"2964287360",
"2608015370",
"2963444790"
],
"abstract": [
"Image-to-image translation tasks have been widely investigated with Generative Adversarial Networks (GANs) and dual learning. However, existing models lack the ability to control the translated results in the target domain and their results usually lack of diversity in the sense that a fixed image usually leads to (almost) deterministic translation result. In this paper, we study a new problem, conditional image-to-image translation, which is to translate an image from the source domain to the target domain conditioned on a given image in the target domain. It requires that the generated image should inherit some domain-specific features of the conditional image from the target domain. Therefore, changing the conditional image in the target domain will lead to diverse translation results for a fixed input image from the source domain, and therefore the conditional input image helps to control the translation results. We tackle this problem with unpaired data based on GANs and dual learning. We twist two conditional translation models (one translation from A domain to B domain, and the other one from B domain to A domain) together for inputs combination and reconstruction while preserving domain independent features. We carry out experiments on men's faces from-to women's faces translation and edges to shoes&bags translations. The results demonstrate the effectiveness of our proposed method.",
"Conditional Generative Adversarial Networks (GANs) for cross-domain image-to-image translation have made much progress recently. Depending on the task complexity, thousands to millions of labeled image pairs are needed to train a conditional GAN. However, human labeling is expensive, even impractical, and large quantities of data may not always be available. Inspired by dual learning from natural language translation, we develop a novel dual-GAN mechanism, which enables image translators to be trained from two sets of unlabeled images from two domains. In our architecture, the primal GAN learns to translate images from domain U to those in domain V, while the dual GAN learns to invert the task. The closed loop made by the primal and dual tasks allows images from either domain to be translated and then reconstructed. Hence a loss function that accounts for the reconstruction error of images can be used to train the translators. Experiments on multiple image translation tasks with unlabeled data show considerable performance gain of DualGAN over a single GAN. For some tasks, DualGAN can even achieve comparable or slightly better results than conditional GAN trained on fully labeled data.",
"Conditional Generative Adversarial Networks (GANs) for cross-domain image-to-image translation have made much progress recently [7, 8, 21, 12, 4, 18]. Depending on the task complexity, thousands to millions of labeled image pairs are needed to train a conditional GAN. However, human labeling is expensive, even impractical, and large quantities of data may not always be available. Inspired by dual learning from natural language translation [23], we develop a novel dual-GAN mechanism, which enables image translators to be trained from two sets of unlabeled images from two domains. In our architecture, the primal GAN learns to translate images from domain U to those in domain V, while the dual GAN learns to invert the task. The closed loop made by the primal and dual tasks allows images from either domain to be translated and then reconstructed. Hence a loss function that accounts for the reconstruction error of images can be used to train the translators. Experiments on multiple image translation tasks with unlabeled data show considerable performance gain of DualGAN over a single GAN. For some tasks, DualGAN can even achieve comparable or slightly better results than conditional GAN trained on fully labeled data."
]
} |
1901.11382 | 2950001871 | In the big data era, the impetus to digitize the vast reservoirs of data trapped in unstructured scanned documents such as invoices, bank documents and courier receipts has gained fresh momentum. The scanning process often results in the introduction of artifacts such as background noise, blur due to camera motion, watermarkings, coffee stains, or faded text. These artifacts pose many readability challenges to current text recognition algorithms and significantly degrade their performance. Existing learning based denoising techniques require a dataset comprising of noisy documents paired with cleaned versions. In such scenarios, a model can be trained to generate clean documents from noisy versions. However, very often in the real world such a paired dataset is not available, and all we have for training our denoising model are unpaired sets of noisy and clean images. This paper explores the use of GANs to generate denoised versions of the noisy documents. In particular, where paired information is available, we formulate the problem as an image-to-image translation task i.e, translating a document from noisy domain ( i.e., background noise, blurred, faded, watermarked ) to a target clean document using Generative Adversarial Networks (GAN). However, in the absence of paired images for training, we employed CycleGAN which is known to learn a mapping between the distributions of the noisy images to the denoised images using unpaired data to achieve image-to-image translation for cleaning the noisy documents. We compare the performance of CycleGAN for document cleaning tasks using unpaired images with a Conditional GAN trained on paired data from the same dataset. Experiments were performed on a public document dataset on which different types of noise were artificially induced, results demonstrate that CycleGAN learns a more robust mapping from the space of noisy to clean documents. | Very few attempts have been made in past for removing watermarks from images. Authors in @cite_15 proposed to use image inpainting to recover the original image. However, the method developed by @cite_24 detects the watermark using statistical methods and subsequently, removes it using image inpainting. To the best of our knowledge, we did not find any work on defading of images. | {
"cite_N": [
"@cite_24",
"@cite_15"
],
"mid": [
"2783930095",
"2766850999"
],
"abstract": [
"This paper introduces a technique to remove visible watermark automatically using image inpainting algorithms. The pending images which need watermark re-moval are assumed to have same resolution and watermark region and we will show this assumption is reasonable. Our proposed technique includes two basic step. The first step is detecting the watermark region, we propose a statistical method to detect the watermark region. Thresholding algorithm for segmentation proceeds at the accumulation image which is calculated by accumulation of the gray-scale maps of pending images. The second step is removing the watermark using image inpainting algorithms. Since watermarks are usually with large re-gion areas, an exemplar-based inpainting algorithm through investigating the sparsity of natural image patches is proposed for this step. Experiments were im-plemented in a test image set of 889 images downloaded from a shopping web-site with the resolution of 800∗800 and same watermark regions.",
"Abstract In this paper, we propose two schemes for visible-watermark removal and reversible image recovery. In the first scheme, we consider the scenario for the image generated by a specific visible (not completely reversible) watermarking algorithm (2017). A run-length coding based method is utilized to compress the difference between the preliminary recovered image and original image. After embedding the difference information invisibly and reversibly, the final embedded image can be exactly recovered to its original version after visible-watermark removal, which avoids the problem of overflow and underflow in (2017). In the second scheme, the scenario of visible-watermark removal for the image generated by any visible watermarking algorithms (no matter the sender and the receiver know the algorithms or not) is considered. The scheme can perfectly remove the embedded visible watermark and can also exactly recover original image with the assist of image inpainting technique. In addition, for both two proposed schemes, the invalid user without the knowledge of secret key cannot achieve reversible recovery for original image. Experimental results demonstrate the effectiveness and superiority of our schemes."
]
} |
1901.11461 | 2949409558 | Mesh models are a promising approach for encoding the structure of 3D objects. Current mesh reconstruction systems predict uniformly distributed vertex locations of a predetermined graph through a series of graph convolutions, leading to compromises with respect to performance or resolution. In this paper, we argue that the graph representation of geometric objects allows for additional structure, which should be leveraged for enhanced reconstruction. Thus, we propose a system which properly benefits from the advantages of the geometric structure of graph encoded objects by introducing (1) a graph convolutional update preserving vertex information; (2) an adaptive splitting heuristic allowing detail to emerge; and (3) a training objective operating both on the local surfaces defined by vertices as well as the global structure defined by the mesh. Our proposed method is evaluated on the task of 3D object reconstruction from images with the ShapeNet dataset, where we demonstrate state of the art performance, both visually and numerically, while having far smaller space requirements by generating adaptive meshes | Mesh models have only recently been used in generation and reconstruction tasks due to the challenging nature of their complex definition @cite_23 . Recent mesh approaches rely on graph representations of meshes, and use GCNs @cite_28 to effectively process them. Our work most closely relates to Neural 3D Mesh Renderer @cite_12 and Pixel2Mesh @cite_23 , which use deformations of a generic pre-defined input mesh, generally a sphere, to form 3D structures. Similarly, Atlas-Net @cite_53 uses deformations over a set of primitive square faces to form 3D shapes. Conceptually similar, there exists numerous papers using class-specific input meshes which are deformed with respect to the given input image @cite_51 @cite_33 @cite_29 @cite_24 @cite_10 @cite_44 . While effective, these approaches require prior knowledge on the target class or access to a model repository. | {
"cite_N": [
"@cite_33",
"@cite_28",
"@cite_10",
"@cite_53",
"@cite_29",
"@cite_24",
"@cite_44",
"@cite_23",
"@cite_51",
"@cite_12"
],
"mid": [
"2883993491",
"2773522905",
"2086550580",
"2254644702",
"2796312544",
"2748348302",
"2016772237",
"2883221003",
"2962928839",
"2728326942"
],
"abstract": [
"We present a unified framework tackling two problems: class-specific 3D reconstruction from a single image, and generation of new 3D shape samples. These tasks have received considerable attention recently; however, existing approaches rely on 3D supervision, annotation of 2D images with keypoints or poses, and or training with multiple views of each object instance. Our framework is very general: it can be trained in similar settings to these existing approaches, while also supporting weaker supervision scenarios. Importantly, it can be trained purely from 2D images, without ground-truth pose annotations, and with a single view per instance. We employ meshes as an output representation, instead of voxels used in most prior work. This allows us to exploit shading information during training, which previous 2D-supervised methods cannot. Thus, our method can learn to generate and reconstruct concave object classes. We evaluate our approach on synthetic data in various settings, showing that (i) it learns to disentangle shape from pose; (ii) using shading in the loss improves performance; (iii) our model is comparable or superior to state-of-the-art voxel-based approaches on quantitative metrics, while producing results that are visually more pleasing; (iv) it still performs well when given supervision weaker than in prior works.",
"One challenge that remains open in 3D deep learning is how to efficiently represent 3D data to feed deep networks. Recent works have relied on volumetric or point cloud representations, but such approaches suffer from a number of issues such as computational complexity, unordered data, and lack of finer geometry. This paper demonstrates that a mesh representation (i.e. vertices and faces to form polygonal surfaces) is able to capture fine-grained geometry for 3D reconstruction tasks. A mesh however is also unstructured data similar to point clouds. We address this problem by proposing a learning framework to infer the parameters of a compact mesh representation rather than learning from the mesh itself. This compact representation encodes a mesh using free-form deformation and a sparse linear combination of models allowing us to reconstruct 3D meshes from single images. In contrast to prior work, we do not rely on silhouettes and landmarks to perform 3D reconstruction. We evaluate our method on synthetic and real-world datasets with very promising results. Our framework efficiently reconstructs 3D objects in a low-dimensional way while preserving its important geometrical aspects.",
"We present a new volumetric method for reconstructing watertight triangle meshes from arbitrary, unoriented point clouds. While previous techniques usually reconstruct surfaces as the zero level-set of a signed distance function, our method uses an unsigned distance function and hence does not require any information about the local surface orientation. Our algorithm estimates local surface confidence values within a dilated crust around the input samples. The surface which maximizes the global confidence is then extracted by computing the minimum cut of a weighted spatial graph structure. We present an algorithm, which efficiently converts this cut into a closed, manifold triangle mesh with a minimal number of vertices. The use of an unsigned distance function avoids the topological noise artifacts caused by misalignment of 3D scans, which are common to most volumetric reconstruction techniques. Due to a hierarchical approach our method efficiently produces solid models of low genus even for noisy and highly irregular data containing large holes, without loosing fine details in densely sampled regions. We show several examples for different application settings such as model generation from raw laser-scanned data, image-based 3D reconstruction, and mesh repair.",
"This article presents a novel approach for 3D mesh labeling by using deep Convolutional Neural Networks (CNNs). Many previous methods on 3D mesh labeling achieve impressive performances by using predefined geometric features. However, the generalization abilities of such low-level features, which are heuristically designed to process specific meshes, are often insufficient to handle all types of meshes. To address this problem, we propose to learn a robust mesh representation that can adapt to various 3D meshes by using CNNs. In our approach, CNNs are first trained in a supervised manner by using a large pool of classical geometric features. In the training process, these low-level features are nonlinearly combined and hierarchically compressed to generate a compact and effective representation for each triangle on the mesh. Based on the trained CNNs and the mesh representations, a label vector is initialized for each triangle to indicate its probabilities of belonging to various object parts. Eventually, a graph-based mesh-labeling algorithm is adopted to optimize the labels of triangles by considering the label consistencies. Experimental results on several public benchmarks show that the proposed approach is robust for various 3D meshes, and outperforms state-of-the-art approaches as well as classic learning algorithms in recognizing mesh labels.",
"We propose an end-to-end deep learning architecture that produces a 3D shape in triangular mesh from a single color image. Limited by the nature of deep neural network, previous methods usually represent a 3D shape in volume or point cloud, and it is non-trivial to convert them to the more ready-to-use mesh model. Unlike the existing methods, our network represents 3D mesh in a graph-based convolutional neural network and produces correct geometry by progressively deforming an ellipsoid, leveraging perceptual features extracted from the input image. We adopt a coarse-to-fine strategy to make the whole deformation procedure stable, and define various of mesh related losses to capture properties of different levels to guarantee visually appealing and physically accurate 3D geometry. Extensive experiments show that our method not only qualitatively produces mesh model with better details, but also achieves higher 3D shape estimation accuracy compared to the state-of-the-art.",
"3D reconstruction from a single image is a key problem in multiple applications ranging from robotic manipulation to augmented reality. Prior methods have tackled this problem through generative models which predict 3D reconstructions as voxels or point clouds. However, these methods can be computationally expensive and miss fine details. We introduce a new differentiable layer for 3D data deformation and use it in DeformNet to learn a model for 3D reconstruction-through-deformation. DeformNet takes an image input, searches the nearest shape template from a database, and deforms the template to match the query image. We evaluate our approach on the ShapeNet dataset and show that - (a) the Free-Form Deformation layer is a powerful new building block for Deep Learning models that manipulate 3D data (b) DeformNet uses this FFD layer combined with shape retrieval for smooth and detail-preserving 3D reconstruction of qualitatively plausible point clouds with respect to a single query image (c) compared to other state-of-the-art 3D reconstruction methods, DeformNet quantitatively matches or outperforms their benchmarks by significant margins. For more information, visit: this https URL .",
"The ability to position a small subset of mesh vertices and produce a meaningful overall deformation of the entire mesh is a fundamental task in mesh editing and animation. However, the class of meaningful deformations varies from mesh to mesh and depends on mesh kinematics, which prescribes valid mesh configurations, and a selection mechanism for choosing among them. Drawing an analogy to the traditional use of skeleton-based inverse kinematics for posing skeletons. we define mesh-based inverse kinematics as the problem of finding meaningful mesh deformations that meet specified vertex constraints.Our solution relies on example meshes to indicate the class of meaningful deformations. Each example is represented with a feature vector of deformation gradients that capture the affine transformations which individual triangles undergo relative to a reference pose. To pose a mesh, our algorithm efficiently searches among all meshes with specified vertex positions to find the one that is closest to some pose in a nonlinear span of the example feature vectors. Since the search is not restricted to the span of example shapes, this produces compelling deformations even when the constraints require poses that are different from those observed in the examples. Furthermore, because the span is formed by a nonlinear blend of the example feature vectors, the blending component of our system may also be used independently to pose meshes by specifying blending weights or to compute multi-way morph sequences.",
"Learned 3D representations of human faces are useful for computer vision problems such as 3D face tracking and reconstruction from images, as well as graphics applications such as character generation and animation. Traditional models learn a latent representation of a face using linear subspaces or higher-order tensor generalizations. Due to this linearity, they can not capture extreme deformations and non-linear expressions. To address this, we introduce a versatile model that learns a non-linear representation of a face using spectral convolutions on a mesh surface. We introduce mesh sampling operations that enable a hierarchical mesh representation that captures non-linear variations in shape and expression at multiple scales within the model. In a variational setting, our model samples diverse realistic 3D faces from a multivariate Gaussian distribution. Our training data consists of 20,466 meshes of extreme expressions captured over 12 different subjects. Despite limited training data, our trained model outperforms state-of-the-art face models with 50 lower reconstruction error, while using 75 fewer parameters. We show that, replacing the expression space of an existing state-of-the-art face model with our model, achieves a lower reconstruction error. Our data, model and code are available at http: coma.is.tue.mpg.de .",
"3D geometric contents are becoming increasingly popular. In this paper, we study the problem of analyzing deforming 3D meshes using deep neural networks. Deforming 3D meshes are flexible to represent 3D animation sequences as well as collections of objects of the same category, allowing diverse shapes with large-scale non-linear deformations. We propose a novel framework which we call mesh variational autoencoders (mesh VAE), to explore the probabilistic latent space of 3D surfaces. The framework is easy to train, and requires very few training examples. We also propose an extended model which allows flexibly adjusting the significance of different latent variables by altering the prior distribution. Extensive experiments demonstrate that our general framework is able to learn a reasonable representation for a collection of deformable shapes, and produce competitive results for a variety of applications, including shape generation, shape interpolation, shape space embedding and shape exploration, outperforming state-of-the-art methods.",
"We introduce a generative model of part-segmented 3D objects: the shape variational auto-encoder (ShapeVAE). The ShapeVAE describes a joint distribution over the existence of object parts, the locations of a dense set of surface points, and over surface normals associated with these points. Our model makes use of a deep encoder-decoder architecture that leverages the part-decomposability of 3D objects to embed high-dimensional shape representations and sample novel instances. Given an input collection of part-segmented objects with dense point correspondences the ShapeVAE is capable of synthesizing novel, realistic shapes, and by performing conditional inference enables imputation of missing parts or surface normals. In addition, by generating both points and surface normals, our model allows for the use of powerful surface-reconstruction methods for mesh synthesis. We provide a quantitative evaluation of the ShapeVAE on shape-completion and test-set log-likelihood tasks and demonstrate that the model performs favourably against strong baselines. We demonstrate qualitatively that the ShapeVAE produces plausible shape samples, and that it captures a semantically meaningful shape-embedding. In addition we show that the ShapeVAE facilitates mesh reconstruction by sampling consistent surface normals."
]
} |
1901.11461 | 2949409558 | Mesh models are a promising approach for encoding the structure of 3D objects. Current mesh reconstruction systems predict uniformly distributed vertex locations of a predetermined graph through a series of graph convolutions, leading to compromises with respect to performance or resolution. In this paper, we argue that the graph representation of geometric objects allows for additional structure, which should be leveraged for enhanced reconstruction. Thus, we propose a system which properly benefits from the advantages of the geometric structure of graph encoded objects by introducing (1) a graph convolutional update preserving vertex information; (2) an adaptive splitting heuristic allowing detail to emerge; and (3) a training objective operating both on the local surfaces defined by vertices as well as the global structure defined by the mesh. Our proposed method is evaluated on the task of 3D object reconstruction from images with the ShapeNet dataset, where we demonstrate state of the art performance, both visually and numerically, while having far smaller space requirements by generating adaptive meshes | The great success of convolutional neural networks in numerous image-based tasks @cite_37 @cite_15 @cite_14 @cite_2 @cite_61 has led to increasing efforts to extend deep networks to domains where graph-structured data is ubiquitous. | {
"cite_N": [
"@cite_61",
"@cite_37",
"@cite_14",
"@cite_2",
"@cite_15"
],
"mid": [
"2770717124",
"2791092480",
"2509155366",
"2893251308",
"2607510315"
],
"abstract": [
"This paper introduces a generalization of Convolutional Neural Networks (CNNs) to graphs with irregular linkage structures, especially heterogeneous graphs with typed nodes and schemas. We propose a novel spatial convolution operation to model the key properties of local connectivity and translation invariance, using high-order connection patterns or motifs. We develop a novel deep architecture Motif-CNN that employs an attention model to combine the features extracted from multiple patterns, thus effectively capturing high-order structural and feature information. Our experiments on semi-supervised node classification on real-world social networks and multiple representative heterogeneous graph datasets indicate significant gains of 6-21 over existing graph CNNs and other state-of-the-art techniques.",
"Convolutional neural networks (CNNs) have massively impacted visual recognition in 2D images, and are now ubiquitous in state-of-the-art approaches. CNNs do not easily extend, however, to data that are not represented by regular grids, such as 3D shape meshes or other graph-structured data, to which traditional local convolution operators do not directly apply. To address this problem, we propose a novel graph-convolution operator to establish correspondences between filter weights and graph neighborhoods with arbitrary connectivity. The key novelty of our approach is that these correspondences are dynamically computed from features learned by the network, rather than relying on predefined static coordinates over the graph as in previous work. We obtain excellent experimental results that significantly improve over previous state-of-the-art shape correspondence results. This shows that our approach can learn effective shape representations from raw input coordinates, without relying on shape descriptors.",
"Automatically detecting illustrations is needed for the target system.Deep Convolutional Neural Networks have been successful in computer vision tasks.DCNN with fine-tuning outperformed the other models including handcrafted features. Systems for aggregating illustrations require a function for automatically distinguishing illustrations from photographs as they crawl the network to collect images. A previous attempt to implement this functionality by designing basic features that were deemed useful for classification achieved an accuracy of only about 58 . On the other hand, deep neural networks had been successful in computer vision tasks, and convolutional neural networks (CNNs) had performed good at extracting such useful image features automatically. We evaluated alternative methods to implement this classification functionality with focus on deep neural networks. As the result of experiments, the method that fine-tuned deep convolutional neural network (DCNN) acquired 96.8 accuracy, outperforming the other models including the custom CNN models that were trained from scratch. We conclude that DCNN with fine-tuning is the best method for implementing a function for automatically distinguishing illustrations from photographs.",
"Convolutional Neural Networks (CNN) are very popular in many fields including computer vision, speech recognition, natural language processing, to name a few. Though deep learning leads to groundbreaking performance in these domains, the networks used are very demanding computationally and are far from real-time even on a GPU, which is not power efficient and therefore does not suit low power systems such as mobile devices. To overcome this challenge, some solutions have been proposed for quantizing the weights and activations of these networks, which accelerate the runtime significantly. Yet, this acceleration comes at the cost of a larger error. The method proposed in this work trains quantized neural networks by noise injection and a learned clamping, which improve the accuracy. This leads to state-of-the-art results on various regression and classification tasks, e.g., ImageNet classification with architectures such as ResNet-18 34 50 with low as 3-bit weights and activations. We implement the proposed solution on an FPGA to demonstrate its applicability for low power real-time applications. The implementation of the paper is available at this https URL",
"Convolutional neural networks provide visual features that perform remarkably well in many computer vision applications. However, training these networks requires significant amounts of supervision. This paper introduces a generic framework to train deep networks, end-to-end, with no supervision. We propose to fix a set of target representations, called Noise As Targets (NAT), and to constrain the deep features to align to them. This domain agnostic approach avoids the standard unsupervised learning issues of trivial solutions and collapsing of features. Thanks to a stochastic batch reassignment strategy and a separable square loss function, it scales to millions of images. The proposed approach produces representations that perform on par with state-of-the-art unsupervised methods on ImageNet and Pascal VOC."
]
} |
1901.11461 | 2949409558 | Mesh models are a promising approach for encoding the structure of 3D objects. Current mesh reconstruction systems predict uniformly distributed vertex locations of a predetermined graph through a series of graph convolutions, leading to compromises with respect to performance or resolution. In this paper, we argue that the graph representation of geometric objects allows for additional structure, which should be leveraged for enhanced reconstruction. Thus, we propose a system which properly benefits from the advantages of the geometric structure of graph encoded objects by introducing (1) a graph convolutional update preserving vertex information; (2) an adaptive splitting heuristic allowing detail to emerge; and (3) a training objective operating both on the local surfaces defined by vertices as well as the global structure defined by the mesh. Our proposed method is evaluated on the task of 3D object reconstruction from images with the ShapeNet dataset, where we demonstrate state of the art performance, both visually and numerically, while having far smaller space requirements by generating adaptive meshes | Early attempts to extend neural networks to deal with arbitrarily structured graphs relied on recursive neural networks @cite_17 @cite_64 @cite_34 . Recently, spectral approaches have emerged as an effective alternative which formulates the convolution as an operation on the spectrum of the graph @cite_18 @cite_59 @cite_62 @cite_9 . Methods operating directly on the graph domain have also been presented. proposed to approximate the filters using the Chebyshev polynomials applied on the Laplacian operator. This approximation was further simplified by . Finally, several works have been introduced exploring well-established deep learning ideas and improving previously reported results @cite_8 @cite_5 @cite_6 @cite_11 . | {
"cite_N": [
"@cite_18",
"@cite_64",
"@cite_62",
"@cite_11",
"@cite_8",
"@cite_9",
"@cite_6",
"@cite_59",
"@cite_5",
"@cite_34",
"@cite_17"
],
"mid": [
"2766931946",
"2963084622",
"2618170429",
"2607041014",
"2951471187",
"2949455170",
"2179596243",
"2574952845",
"2788715907",
"2139906443",
"2809418595"
],
"abstract": [
"We propose a generalization of convolutional neural networks (CNNs) to irregular domains, through the use of an inferred graph structure. In more details, we introduce a three-step methodology to create convolutional layers that are adapted to the signals to process: 1) From a training set of signals, infer a graph representing the topology on which they evolve; 2) Identify translation operators in the vertex domain; 3) Emulate a convolution operator by translating a localized kernel on the graph. Using these layers, a convolutional neural network is built, and is trained on the initial signals to perform a classification task. Contributions are twofold. First, we adapt a definition of translations on graphs to make them more robust to irregularities, and to take into account locality of the kernel. Second, we introduce a procedure to build CNNs from data. We apply our methodology on a scrambled version of the CIFAR-10 and Haxby datasets. Without using any knowledge on the signals, we significantly outperform existing methods. Moreover, our approach extends classical CNNs on images in the sense that such networks are a particular case of our approach when the inferred graph is a grid.",
"The rise of graph-structured data such as social networks, regulatory networks, citation graphs, and functional brain networks, in combination with resounding success of deep learning in various applications, has brought the interest in generalizing deep learning models to non-Euclidean domains. In this paper, we introduce a new spectral domain convolutional architecture for deep learning on graphs. The core ingredient of our model is a new class of parametric rational complex functions (Cayley polynomials) allowing to efficiently compute spectral filters on graphs that specialize on frequency bands of interest. Our model generates rich spectral filters that are localized in space, scales linearly with the size of the input data for sparsely connected graphs, and can handle different constructions of Laplacian operators. Extensive experimental results show the superior performance of our approach, in comparison to other spectral domain convolutional architectures, on spectral image classification, community detection, vertex classification, and matrix completion tasks.",
"The rise of graph-structured data such as social networks, regulatory networks, citation graphs, and functional brain networks, in combination with resounding success of deep learning in various applications, has brought the interest in generalizing deep learning models to non-Euclidean domains. In this paper, we introduce a new spectral domain convolutional architecture for deep learning on graphs. The core ingredient of our model is a new class of parametric rational complex functions (Cayley polynomials) allowing to efficiently compute spectral filters on graphs that specialize on frequency bands of interest. Our model generates rich spectral filters that are localized in space, scales linearly with the size of the input data for sparsely-connected graphs, and can handle different constructions of Laplacian operators. Extensive experimental results show the superior performance of our approach, in comparison to other spectral domain convolutional architectures, on spectral image classification, community detection, vertex classification and matrix completion tasks.",
"Convolutional neural networks have recently demonstrated high-quality reconstruction for single-image super-resolution. In this paper, we propose the Laplacian Pyramid Super-Resolution Network (LapSRN) to progressively reconstruct the sub-band residuals of high-resolution images. At each pyramid level, our model takes coarse-resolution feature maps as input, predicts the high-frequency residuals, and uses transposed convolutions for upsampling to the finer level. Our method does not require the bicubic interpolation as the pre-processing step and thus dramatically reduces the computational complexity. We train the proposed LapSRN with deep supervision using a robust Charbonnier loss function and achieve high-quality reconstruction. Furthermore, our network generates multi-scale predictions in one feed-forward pass through the progressive reconstruction, thereby facilitates resource-aware applications. Extensive quantitative and qualitative evaluations on benchmark datasets show that the proposed algorithm performs favorably against the state-of-the-art methods in terms of speed and accuracy.",
"Convolutional neural networks have recently demonstrated high-quality reconstruction for single-image super-resolution. In this paper, we propose the Laplacian Pyramid Super-Resolution Network (LapSRN) to progressively reconstruct the sub-band residuals of high-resolution images. At each pyramid level, our model takes coarse-resolution feature maps as input, predicts the high-frequency residuals, and uses transposed convolutions for upsampling to the finer level. Our method does not require the bicubic interpolation as the pre-processing step and thus dramatically reduces the computational complexity. We train the proposed LapSRN with deep supervision using a robust Charbonnier loss function and achieve high-quality reconstruction. Furthermore, our network generates multi-scale predictions in one feed-forward pass through the progressive reconstruction, thereby facilitates resource-aware applications. Extensive quantitative and qualitative evaluations on benchmark datasets show that the proposed algorithm performs favorably against the state-of-the-art methods in terms of speed and accuracy.",
"A number of problems can be formulated as prediction on graph-structured data. In this work, we generalize the convolution operator from regular grids to arbitrary graphs while avoiding the spectral domain, which allows us to handle graphs of varying size and connectivity. To move beyond a simple diffusion, filter weights are conditioned on the specific edge labels in the neighborhood of a vertex. Together with the proper choice of graph coarsening, we explore constructing deep neural networks for graph classification. In particular, we demonstrate the generality of our formulation in point cloud classification, where we set the new state of the art, and on a graph classification dataset, where we outperform other deep learning approaches. The source code is available at this https URL",
"We propose a new method for creating computationally efficient convolutional neural networks (CNNs) by using low-rank representations of convolutional filters. Rather than approximating filters in previously-trained networks with more efficient versions, we learn a set of small basis filters from scratch; during training, the network learns to combine these basis filters into more complex filters that are discriminative for image classification. To train such networks, a novel weight initialization scheme is used. This allows effective initialization of connection weights in convolutional layers composed of groups of differently-shaped filters. We validate our approach by applying it to several existing CNN architectures and training these networks from scratch using the CIFAR, ILSVRC and MIT Places datasets. Our results show similar or higher accuracy than conventional CNNs with much less compute. Applying our method to an improved version of VGG-11 network using global max-pooling, we achieve comparable validation accuracy using 41 less compute and only 24 of the original VGG-11 model parameters; another variant of our method gives a 1 percentage point increase in accuracy over our improved VGG-11 model, giving a top-5 center-crop validation accuracy of 89.7 while reducing computation by 16 relative to the original VGG-11 model. Applying our method to the GoogLeNet architecture for ILSVRC, we achieved comparable accuracy with 26 less compute and 41 fewer model parameters. Applying our method to a near state-of-the-art network for CIFAR, we achieved comparable accuracy with 46 less compute and 55 fewer parameters.",
"In this paper, we propose a novel deep convolutional neural network (CNN)-based algorithm for solving ill-posed inverse problems. Regularized iterative algorithms have emerged as the standard approach to ill-posed inverse problems in the past few decades. These methods produce excellent results, but can be challenging to deploy in practice due to factors including the high computational cost of the forward and adjoint operators and the difficulty of hyperparameter selection. The starting point of this paper is the observation that unrolled iterative methods have the form of a CNN (filtering followed by pointwise non-linearity) when the normal operator ( @math , where @math is the adjoint of the forward imaging operator, @math ) of the forward model is a convolution. Based on this observation, we propose using direct inversion followed by a CNN to solve normal-convolutional inverse problems. The direct inversion encapsulates the physical model of the system, but leads to artifacts when the problem is ill posed; the CNN combines multiresolution decomposition and residual learning in order to learn to remove these artifacts while preserving image structure. We demonstrate the performance of the proposed network in sparse-view reconstruction (down to 50 views) on parallel beam X-ray computed tomography in synthetic phantoms as well as in real experimental sinograms. The proposed network outperforms total variation-regularized iterative reconstruction for the more realistic phantoms and requires less than a second to reconstruct a @math image on the GPU.",
"In recent years considerable research efforts have been devoted to compression techniques of convolutional neural networks (CNNs). Many works so far have focused on CNN connection pruning methods which produce sparse parameter tensors in convolutional or fully-connected layers. It has been demonstrated in several studies that even simple methods can effectively eliminate connections of a CNN. However, since these methods make parameter tensors just sparser but no smaller, the compression may not transfer directly to acceleration without support from specially designed hardware. In this paper, we propose an iterative approach named Auto-balanced Filter Pruning, where we pre-train the network in an innovative auto-balanced way to transfer the representational capacity of its convolutional layers to a fraction of the filters, prune the redundant ones, then re-train it to restore the accuracy. In this way, a smaller version of the original network is learned and the floating-point operations (FLOPs) are reduced. By applying this method on several common CNNs, we show that a large portion of the filters can be discarded without obvious accuracy drop, leading to significant reduction of computational burdens. Concretely, we reduce the inference cost of LeNet-5 on MNIST, VGG-16 and ResNet-56 on CIFAR-10 by 95.1 , 79.7 and 60.9 , respectively. Copyright © 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.",
"This paper presents a new approach for learning in structured domains (SDs) using a constructive neural network for graphs (NN4G). The new model allows the extension of the input domain for supervised neural networks to a general class of graphs including both acyclic cyclic, directed undirected labeled graphs. In particular, the model can realize adaptive contextual transductions, learning the mapping from graphs for both classification and regression tasks. In contrast to previous neural networks for structures that had a recursive dynamics, NN4G is based on a constructive feedforward architecture with state variables that uses neurons with no feedback connections. The neurons are applied to the input graphs by a general traversal process that relaxes the constraints of previous approaches derived by the causality assumption over hierarchical input data. Moreover, the incremental approach eliminates the need to introduce cyclic dependencies in the definition of the system state variables. In the traversal process, the NN4G units exploit (local) contextual information of the graphs vertices. In spite of the simplicity of the approach, we show that, through the compositionality of the contextual information developed by the learning, the model can deal with contextual information that is incrementally extended according to the graphs topology. The effectiveness and the generality of the new approach are investigated by analyzing its theoretical properties and providing experimental results.",
"Convolutional neural networks (CNNs) have achieved great success on grid-like data such as images, but face tremendous challenges in learning from more generic data such as graphs. In CNNs, the trainable local filters enable the automatic extraction of high-level features. The computation with filters requires a fixed number of ordered units in the receptive fields. However, the number of neighboring units is neither fixed nor are they ordered in generic graphs, thereby hindering the applications of convolutional operations. Here, we address these challenges by proposing the learnable graph convolutional layer (LGCL). LGCL automatically selects a fixed number of neighboring nodes for each feature based on value ranking in order to transform graph data into grid-like structures in 1-D format, thereby enabling the use of regular convolutional operations on generic graphs. To enable model training on large-scale graphs, we propose a sub-graph training method to reduce the excessive memory and computational resource requirements suffered by prior methods on graph convolutions. Our experimental results on node classification tasks in both transductive and inductive learning settings demonstrate that our methods can achieve consistently better performance on the Cora, Citeseer, Pubmed citation network, and protein-protein interaction network datasets. Our results also indicate that the proposed methods using sub-graph training strategy are more efficient as compared to prior approaches."
]
} |
1901.10997 | 2913686213 | Many long short-term memory (LSTM) applications need fast yet compact models. Neural network compression approaches, such as the grow-and-prune paradigm, have proved to be promising for cutting down network complexity by skipping insignificant weights. However, current compression strategies are mostly hardware-agnostic and network complexity reduction does not always translate into execution efficiency. In this work, we propose a hardware-guided symbiotic training methodology for compact, accurate, yet execution-efficient inference models. It is based on our observation that hardware may introduce substantial non-monotonic behavior, which we call the latency hysteresis effect, when evaluating network size vs. inference latency. This observation raises question about the mainstream smaller-dimension-is-better compression strategy, which often leads to a sub-optimal model architecture. By leveraging the hardware-impacted hysteresis effect and sparsity, we are able to achieve the symbiosis of model compactness and accuracy with execution efficiency, thus reducing LSTM latency while increasing its accuracy. We have evaluated our algorithms on language modeling and speech recognition applications. Relative to the traditional stacked LSTM architecture obtained for the Penn Treebank dataset, we reduce the number of parameters by 18.0x (30.5x) and measured run-time latency by up to 2.4x (5.2x) on Nvidia GPUs (Intel Xeon CPUs) without any accuracy degradation. For the DeepSpeech2 architecture obtained for the AN4 dataset, we reduce the number of parameters by 7.0x (19.4x), word error rate from 12.9 to 9.9 (10.4 ), and measured run-time latency by up to 1.7x (2.4x) on Nvidia GPUs (Intel Xeon CPUs). Thus, our method yields compact, accurate, yet execution-efficient inference models. | Various attempts have been made to improve the efficiency of LSTM models. One direction focuses on improving the LSTM cells. The gated recurrent unit (GRU) utilizes reset and update gates to achieve a similar performance to an LSTM while reducing computational cost @cite_5 . Quasi-RNN explores the intrinsic parallelism of time series data to outperform an LSTM for the same hidden state width @cite_14 . H-LSTM incorporates deeper control gates to reduce the number of external stacked layers. It achieves higher accuracy than the GRU and LSTM with fewer parameters @cite_19 . | {
"cite_N": [
"@cite_19",
"@cite_5",
"@cite_14"
],
"mid": [
"2953188482",
"2594230714",
"2319453305"
],
"abstract": [
"Recurrent neural networks (RNNs) have shown clear superiority in sequence modeling, particularly the ones with gated units, such as long short-term memory (LSTM) and gated recurrent unit (GRU). However, the dynamic properties behind the remarkable performance remain unclear in many applications, e.g., automatic speech recognition (ASR). This paper employs visualization techniques to study the behavior of LSTM and GRU when performing speech recognition tasks. Our experiments show some interesting patterns in the gated memory, and some of them have inspired simple yet effective modifications on the network structure. We report two of such modifications: (1) lazy cell update in LSTM, and (2) shortcut connections for residual learning. Both modifications lead to more comprehensible and powerful networks.",
"Sophisticated gated recurrent neural network architectures like LSTMs and GRUs have been shown to be highly effective in a myriad of applications. We develop an un-gated unit, the statistical recurrent unit (SRU), that is able to learn long term dependencies in data by only keeping moving averages of statistics. The SRU's architecture is simple, un-gated, and contains a comparable number of parameters to LSTMs; yet, SRUs perform favorably to more sophisticated LSTM and GRU alternatives, often outperforming one or both in various tasks. We show the efficacy of SRUs as compared to LSTMs and GRUs in an unbiased manner by optimizing respective architectures' hyperparameters for both synthetic and real-world tasks.",
"Recurrent neural networks (RNN) have been very successful in handling sequence data. However, understanding RNN and finding the best practices for RNN learning is a difficult task, partly because there are many competing and complex hidden units, such as the long short-term memory (LSTM) and the gated recurrent unit (GRU). We propose a gated unit for RNN, named as minimal gated unit (MGU), since it only contains one gate, which is a minimal design among all gated hidden units. The design of MGU benefits from evaluation results on LSTM and GRU in the literature. Experiments on various sequence data show that MGU has comparable accuracy with GRU, but has a simpler structure, fewer parameters, and faster training. Hence, MGU is suitable in RNN's applications. Its simple architecture also means that it is easier to evaluate and tune, and in principle it is easier to study MGU's properties theoretically and empirically."
]
} |
1901.10997 | 2913686213 | Many long short-term memory (LSTM) applications need fast yet compact models. Neural network compression approaches, such as the grow-and-prune paradigm, have proved to be promising for cutting down network complexity by skipping insignificant weights. However, current compression strategies are mostly hardware-agnostic and network complexity reduction does not always translate into execution efficiency. In this work, we propose a hardware-guided symbiotic training methodology for compact, accurate, yet execution-efficient inference models. It is based on our observation that hardware may introduce substantial non-monotonic behavior, which we call the latency hysteresis effect, when evaluating network size vs. inference latency. This observation raises question about the mainstream smaller-dimension-is-better compression strategy, which often leads to a sub-optimal model architecture. By leveraging the hardware-impacted hysteresis effect and sparsity, we are able to achieve the symbiosis of model compactness and accuracy with execution efficiency, thus reducing LSTM latency while increasing its accuracy. We have evaluated our algorithms on language modeling and speech recognition applications. Relative to the traditional stacked LSTM architecture obtained for the Penn Treebank dataset, we reduce the number of parameters by 18.0x (30.5x) and measured run-time latency by up to 2.4x (5.2x) on Nvidia GPUs (Intel Xeon CPUs) without any accuracy degradation. For the DeepSpeech2 architecture obtained for the AN4 dataset, we reduce the number of parameters by 7.0x (19.4x), word error rate from 12.9 to 9.9 (10.4 ), and measured run-time latency by up to 1.7x (2.4x) on Nvidia GPUs (Intel Xeon CPUs). Thus, our method yields compact, accurate, yet execution-efficient inference models. | Network compression techniques, such as the grow-and-prune paradigm, have recently emerged as another direction for reducing LSTM redundancy. The pruning method was initially shown to be effective on large CNNs by demonstrating the reduction in the number of parameters in AlexNet by 9 @math and VGG by 13 @math for the well-known ImageNet dataset, without any accuracy loss @cite_37 . Follow-up works have successfully scaled this technique to LSTMs @cite_40 @cite_0 @cite_38 . For example, a recent work proposes structured pruning for LSTMs through group LASSO regularization @cite_0 . Network growth is a complementary method to pruning. It enables a more sparse yet accurate model to be obtained before pruning starts @cite_31 . A grow-and-prune paradigm typically reduces the number of parameters in CNNs @cite_31 and LSTMs @cite_19 by another 2 @math . However, all these methods are hardware-agnostic. Most of them utilize monotonic optimization metrics, e.g., smaller matrix dimensions or fewer multiply-accumulate operations, hence optimize towards slimmer or more sparse models that may not necessarily translate into execution efficiency. | {
"cite_N": [
"@cite_38",
"@cite_37",
"@cite_0",
"@cite_19",
"@cite_40",
"@cite_31"
],
"mid": [
"2788715907",
"2515385951",
"2962965870",
"2735189290",
"2952344559",
"2884180697"
],
"abstract": [
"In recent years considerable research efforts have been devoted to compression techniques of convolutional neural networks (CNNs). Many works so far have focused on CNN connection pruning methods which produce sparse parameter tensors in convolutional or fully-connected layers. It has been demonstrated in several studies that even simple methods can effectively eliminate connections of a CNN. However, since these methods make parameter tensors just sparser but no smaller, the compression may not transfer directly to acceleration without support from specially designed hardware. In this paper, we propose an iterative approach named Auto-balanced Filter Pruning, where we pre-train the network in an innovative auto-balanced way to transfer the representational capacity of its convolutional layers to a fraction of the filters, prune the redundant ones, then re-train it to restore the accuracy. In this way, a smaller version of the original network is learned and the floating-point operations (FLOPs) are reduced. By applying this method on several common CNNs, we show that a large portion of the filters can be discarded without obvious accuracy drop, leading to significant reduction of computational burdens. Concretely, we reduce the inference cost of LeNet-5 on MNIST, VGG-16 and ResNet-56 on CIFAR-10 by 95.1 , 79.7 and 60.9 , respectively. Copyright © 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.",
"The success of CNNs in various applications is accompanied by a significant increase in the computation and parameter storage costs. Recent efforts toward reducing these overheads involve pruning and compressing the weights of various layers without hurting original accuracy. However, magnitude-based pruning of weights reduces a significant number of parameters from the fully connected layers and may not adequately reduce the computation costs in the convolutional layers due to irregular sparsity in the pruned networks. We present an acceleration method for CNNs, where we prune filters from CNNs that are identified as having a small effect on the output accuracy. By removing whole filters in the network together with their connecting feature maps, the computation costs are reduced significantly. In contrast to pruning weights, this approach does not result in sparse connectivity patterns. Hence, it does not need the support of sparse convolution libraries and can work with existing efficient BLAS libraries for dense matrix multiplications. We show that even simple filter pruning techniques can reduce inference costs for VGG-16 by up to 34 and ResNet-110 by up to 38 on CIFAR10 while regaining close to the original accuracy by retraining the networks.",
"The success of CNNs in various applications is accompanied by a significant increase in the computation and parameter storage costs. Recent efforts toward reducing these overheads involve pruning and compressing the weights of various layers without hurting original accuracy. However, magnitude-based pruning of weights reduces a significant number of parameters from the fully connected layers and may not adequately reduce the computation costs in the convolutional layers due to irregular sparsity in the pruned networks. We present an acceleration method for CNNs, where we prune filters from CNNs that are identified as having a small effect on the output accuracy. By removing whole filters in the network together with their connecting feature maps, the computation costs are reduced significantly. In contrast to pruning weights, this approach does not result in sparse connectivity patterns. Hence, it does not need the support of sparse convolution libraries and can work with existing efficient BLAS libraries for dense matrix multiplications. We show that even simple filter pruning techniques can reduce inference costs for VGG-16 by up to 34 and ResNet-110 by up to 38 on CIFAR10 while regaining close to the original accuracy by retraining the networks.",
"Network pruning is an effective way to accelerate Convolutional Neural Networks (CNNs). In recent years, structured pruning methods are proposed in favor of unstructured methods as they have shown greater speedup in practical use. Existing structured methods does pruning along two main dimensions: 3D-filter wise, i.e., remove a 3D-fllter as a whole, and filter-shape wise, i.e., remove a same position from all 3D-filters. In this work, we propose a new group-wise 2D-fllter pruning approach that is orthogonal and complementary to the existing methods. The proposed approach removes a portion of 2D-fllters from each 3D-filter according to the pruning patterns learned from the data, and leads to compressed models that do not require sophisticated implementation of convolution operations. A fine-tuning process is followed to recover the accuracy. The knowledge distillation (KD) framework is explored in the fine-tuning process to improve the performance. We present our method for learning the pruning pattens as well as the fine-tuning strategy based on knowledge distillation. The proposed approach is validated on two representative CNN models — ZF and VGG16, pre-trained on ILSVRC12. Experimental results demonstrate the effectiveness of our approach. In VGG16, we get even higher accuracy after speeding-up the network by 4 times.",
"Model pruning seeks to induce sparsity in a deep neural network's various connection matrices, thereby reducing the number of nonzero-valued parameters in the model. Recent reports (, 2015; , 2017) prune deep networks at the cost of only a marginal loss in accuracy and achieve a sizable reduction in model size. This hints at the possibility that the baseline models in these experiments are perhaps severely over-parameterized at the outset and a viable alternative for model compression might be to simply reduce the number of hidden units while maintaining the model's dense connection structure, exposing a similar trade-off in model size and accuracy. We investigate these two distinct paths for model compression within the context of energy-efficient inference in resource-constrained environments and propose a new gradual pruning technique that is simple and straightforward to apply across a variety of models datasets with minimal tuning and can be seamlessly incorporated within the training process. We compare the accuracy of large, but pruned models (large-sparse) and their smaller, but dense (small-dense) counterparts with identical memory footprint. Across a broad range of neural network architectures (deep CNNs, stacked LSTM, and seq2seq LSTM models), we find large-sparse models to consistently outperform small-dense models and achieve up to 10x reduction in number of non-zero parameters with minimal loss in accuracy.",
"Weight pruning methods of deep neural networks have been demonstrated to achieve a good model pruning ratio without loss of accuracy, thereby alleviating the significant computation storage requirements of large-scale DNNs. Structured weight pruning methods have been proposed to overcome the limitation of irregular network structure and demonstrated actual GPU acceleration. However, the pruning ratio and GPU acceleration are limited when accuracy needs to be maintained. In this work, we overcome pruning ratio and GPU acceleration limitations by proposing a unified, systematic framework of structured weight pruning for DNNs, named ADAM-ADMM. It is a framework that can be used to induce different types of structured sparsity, such as filter-wise, channel-wise, and shape-wise sparsity, as well non-structured sparsity. The proposed framework incorporates stochastic gradient descent with ADMM, and can be understood as a dynamic regularization method in which the regularization target is analytically updated in each iteration. A significant improvement in structured weight pruning ratio is achieved without loss of accuracy, along with fast convergence rate. With a small sparsity degree of 33.3 on the convolutional layers, we achieve 1.64 accuracy enhancement for the AlexNet model. This is obtained by mitigation of overfitting. Without loss of accuracy on the AlexNet model, we achieve 2.58x and 3.65x average measured speedup on two GPUs, clearly outperforming the prior work. The average speedups reach 2.77x and 7.5x when allowing a moderate accuracy loss of 2 . In this case the model compression for convolutional layers is 13.2x, corresponding to 10.5x CPU speedup. Our experiments on ResNet model and on other datasets like UCF101 and CIFAR-10 demonstrate the consistently higher performance of our framework. Our models and codes are released at this https URL"
]
} |
1901.11168 | 2972498781 | The emergence of continuous health monitoring and the availability of an enormous amount of time series data has provided a great opportunity for the advancement of personal health tracking. In recent years, unsupervised learning methods have drawn special attention of researchers to tackle the sparse annotation of health data and real-time detection of anomalies has been a central problem of interest. However, one problem that has not been well addressed before is the early prediction of forthcoming negative health events. Early signs of an event can introduce subtle and gradual changes in the health signal prior to its onset, detection of which can be invaluable in effective prevention. In this study, we first demonstrate our observations on the shortcoming of widely adopted anomaly detection methods in uncovering the changes prior to a negative health event. We then propose a framework which relies on online clustering of signal segment representations which are automatically learned by a specially designed LSTM auto-encoder. We benchmark our results on the publicly available MIT-PICS dataset and show the effectiveness of our approach by predicting Bradycardia events in infants 1.3 minutes ahead of time with 68 AUC score on average, with no label supervision. Results of our study can indicate the viability of our approach in the early detection of health events in other applications as well. | With the rise of unsupervised deep learning models especially auto-encoders @cite_19 and their great performance in other domains such as image recognition @cite_1 , their application has recently emerged in wireless health for detection of anomalies in health signals such as ECG signals. @cite_7 @cite_14 are among studies that employed auto-encoders on ECG to distinguish anomalous parts from the healthy ones. For this aim, the reconstruction error from the auto-encoder that is trained on normal data is tracked to find sudden jumps, motivated by the idea that such a model cannot reconstruct anomalous intervals of data accurately. Auto-encoders have successfully replaced prior approaches such as classifiers @cite_6 which require large annotated datasets, alongside statistical clustering models @cite_4 , and future value predictor models @cite_12 that both are not easily generalizable to other applications. | {
"cite_N": [
"@cite_14",
"@cite_4",
"@cite_7",
"@cite_1",
"@cite_6",
"@cite_19",
"@cite_12"
],
"mid": [
"2289846183",
"2922995320",
"2909671697",
"2743138268",
"2558464646",
"2474046725",
"2150150736"
],
"abstract": [
"In this paper, we propose a novel approach based on deep learning for active classification of electrocardiogram (ECG) signals. To this end, we learn a suitable feature representation from the raw ECG data in an unsupervised way using stacked denoising autoencoders (SDAEs) with sparsity constraint. After this feature learning phase, we add a softmax regression layer on the top of the resulting hidden representation layer yielding the so-called deep neural network (DNN). During the interaction phase, we allow the expert at each iteration to label the most relevant and uncertain ECG beats in the test record, which are then used for updating the DNN weights. As ranking criteria, the method relies on the DNN posterior probabilities to associate confidence measures such as entropy and Breaking-Ties (BT) to each test beat in the ECG record under analysis. In the experiments, we validate the method on the well-known MIT-BIH arrhythmia database as well as two other databases called INCART, and SVDB, respectively. Furthermore, we follow the recommendations of the Association for the Advancement of Medical Instrumentation (AAMI) for class labeling and results presentation. The results obtained show that the newly proposed approach provides significant accuracy improvements with less expert interaction and faster online retraining compared to state-of-the-art methods.",
"We propose the Autoencoding Binary Classifiers (ABC), a novel supervised anomaly detector based on the Autoencoder (AE). There are two main approaches in anomaly detection: supervised and unsupervised. The supervised approach accurately detects the known anomalies included in training data, but it cannot detect the unknown anomalies. Meanwhile, the unsupervised approach can detect both known and unknown anomalies that are located away from normal data points. However, it does not detect known anomalies as accurately as the supervised approach. Furthermore, even if we have labeled normal data points and anomalies, the unsupervised approach cannot utilize these labels. The ABC is a probabilistic binary classifier that effectively exploits the label information, where normal data points are modeled using the AE as a component. By maximizing the likelihood, the AE in the proposed ABC is trained to minimize the reconstruction error for normal data points, and to maximize it for known anomalies. Since our approach becomes able to reconstruct the normal data points accurately and fails to reconstruct the known and unknown anomalies, it can accurately discriminate both known and unknown anomalies from normal data points. Experimental results show that the ABC achieves higher detection performance than existing supervised and unsupervised methods.",
"The success of deep neural networks often relies on a large amount of labeled examples, which can be difficult to obtain in many real scenarios. To address this challenge, unsupervised methods are strongly preferred for training neural networks without using any labeled data. In this paper, we present a novel paradigm of unsupervised representation learning by Auto-Encoding Transformation (AET) in contrast to the conventional Auto-Encoding Data (AED) approach. Given a randomly sampled transformation, AET seeks to predict it merely from the encoded features as accurately as possible at the output end. The idea is the following: as long as the unsupervised features successfully encode the essential information about the visual structures of original and transformed images, the transformation can be well predicted. We will show that this AET paradigm allows us to instantiate a large variety of transformations, from parameterized, to non-parameterized and GAN-induced ones. Our experiments show that AET greatly improves over existing unsupervised approaches, setting new state-of-the-art performances being greatly closer to the upper bounds by their fully supervised counterparts on CIFAR-10, ImageNet and Places datasets.",
"Deep autoencoders, and other deep neural networks, have demonstrated their effectiveness in discovering non-linear features across many problem domains. However, in many real-world problems, large outliers and pervasive noise are commonplace, and one may not have access to clean training data as required by standard deep denoising autoencoders. Herein, we demonstrate novel extensions to deep autoencoders which not only maintain a deep autoencoders' ability to discover high quality, non-linear features but can also eliminate outliers and noise without access to any clean training data. Our model is inspired by Robust Principal Component Analysis, and we split the input data X into two parts, @math , where @math can be effectively reconstructed by a deep autoencoder and @math contains the outliers and noise in the original data X. Since such splitting increases the robustness of standard deep autoencoders, we name our model a \"Robust Deep Autoencoder (RDA)\". Further, we present generalizations of our results to grouped sparsity norms which allow one to distinguish random anomalies from other types of structured corruptions, such as a collection of features being corrupted across many instances or a collection of instances having more corruptions than their fellows. Such \"Group Robust Deep Autoencoders (GRDA)\" give rise to novel anomaly detection approaches whose superior performance we demonstrate on a selection of benchmark problems.",
"The work explores the use of denoising autoencoders (DAEs) for brain lesion detection, segmentation, and false-positive reduction. Stacked denoising autoencoders (SDAEs) were pretrained using a large number of unlabeled patient volumes and fine-tuned with patches drawn from a limited number of patients (n=20, 40, 65). The results show negligible loss in performance even when SDAE was fine-tuned using 20 labeled patients. Low grade glioma (LGG) segmentation was achieved using a transfer learning approach in which a network pretrained with high grade glioma data was fine-tuned using LGG image patches. The networks were also shown to generalize well and provide good segmentation on unseen BraTS 2013 and BraTS 2015 test data. The manuscript also includes the use of a single layer DAE, referred to as novelty detector (ND). ND was trained to accurately reconstruct nonlesion patches. The reconstruction error maps of test data were used to localize lesions. The error maps were shown to assign unique error distributions to various constituents of the glioma, enabling localization. The ND learns the nonlesion brain accurately as it was also shown to provide good segmentation performance on ischemic brain lesions in images from a different database.",
"Mechanical devices such as engines, vehicles, aircrafts, etc., are typically instrumented with numerous sensors to capture the behavior and health of the machine. However, there are often external factors or variables which are not captured by sensors leading to time-series which are inherently unpredictable. For instance, manual controls and or unmonitored environmental conditions or load may lead to inherently unpredictable time-series. Detecting anomalies in such scenarios becomes challenging using standard approaches based on mathematical models that rely on stationarity, or prediction models that utilize prediction errors to detect anomalies. We propose a Long Short Term Memory Networks based Encoder-Decoder scheme for Anomaly Detection (EncDec-AD) that learns to reconstruct 'normal' time-series behavior, and thereafter uses reconstruction error to detect anomalies. We experiment with three publicly available quasi predictable time-series datasets: power demand, space shuttle, and ECG, and two real-world engine datasets with both predictive and unpredictable behavior. We show that EncDec-AD is robust and can detect anomalies from predictable, unpredictable, periodic, aperiodic, and quasi-periodic time-series. Further, we show that EncDec-AD is able to detect anomalies from short time-series (length as small as 30) as well as long time-series (length as large as 500).",
"Signals from peripheral physiology (e.g., ECG, EMG, and GSR) in conjunction with machine learning techniques can be used for the automatic detection of affective states. The affect detector can be user-independent, where it is expected to generalize to novel users, or user-dependent, where it is tailored to a specific user. Previous studies have reported some success in detecting affect from physiological signals, but much of the work has focused on induced affect or acted expressions instead of contextually constrained spontaneous expressions of affect. This study addresses these issues by developing and evaluating user-independent and user-dependent physiology-based detectors of nonbasic affective states (e.g., boredom, confusion, curiosity) that were trained and validated on naturalistic data collected during interactions between 27 students and AutoTutor, an intelligent tutoring system with conversational dialogues. There is also no consensus on which techniques (i.e., feature selection or classification methods) work best for this type of data. Therefore, this study also evaluates the efficacy of affect detection using a host of feature selection and classification techniques on three physiological signals (ECG, EMG, and GSR) and their combinations. Two feature selection methods and nine classifiers were applied to the problem of recognizing eight affective states (boredom, confusion, curiosity, delight, flow -engagement, surprise, and neutral). The results indicated that the user-independent modeling approach was not feasible; however, a mean kappa score of 0.25 was obtained for user-dependent models that discriminated among the most frequent emotions. The results also indicated that k-nearest neighbor and Linear Bayes Normal Classifier (LBNC) classifiers yielded the best affect detection rates. Single channel ECG, EMG, and GSR and three-channel multimodal models were generally more diagnostic than two--channel models."
]
} |
1901.11168 | 2972498781 | The emergence of continuous health monitoring and the availability of an enormous amount of time series data has provided a great opportunity for the advancement of personal health tracking. In recent years, unsupervised learning methods have drawn special attention of researchers to tackle the sparse annotation of health data and real-time detection of anomalies has been a central problem of interest. However, one problem that has not been well addressed before is the early prediction of forthcoming negative health events. Early signs of an event can introduce subtle and gradual changes in the health signal prior to its onset, detection of which can be invaluable in effective prevention. In this study, we first demonstrate our observations on the shortcoming of widely adopted anomaly detection methods in uncovering the changes prior to a negative health event. We then propose a framework which relies on online clustering of signal segment representations which are automatically learned by a specially designed LSTM auto-encoder. We benchmark our results on the publicly available MIT-PICS dataset and show the effectiveness of our approach by predicting Bradycardia events in infants 1.3 minutes ahead of time with 68 AUC score on average, with no label supervision. Results of our study can indicate the viability of our approach in the early detection of health events in other applications as well. | LSTM auto-encoders @cite_0 were later introduced in learning representations of videos and improved feature extraction by capturing temporal features of the signal. They were later used in time series analysis as well @cite_17 . Moreover, Two recent studies have shown improved performance of auto-encoders in more complex anomaly detection settings by utilizing the encoded representation from auto-encoders in offline clustering of anomalies @cite_5 or detection of signal change point by comparing neighbor segment representations @cite_13 . Although these studies follow different goals, we employ their finding in this study in building our model. | {
"cite_N": [
"@cite_0",
"@cite_5",
"@cite_13",
"@cite_17"
],
"mid": [
"2952694903",
"2175030374",
"2952453038",
"2116435618"
],
"abstract": [
"With the growing popularity of short-form video sharing platforms such as Instagram and Vine , there has been an increasing need for techniques that automatically extract highlights from video. Whereas prior works have approached this problem with heuristic rules or supervised learning, we present an unsupervised learning approach that takes advantage of the abundance of user-edited videos on social media websites such as YouTube. Based on the idea that the most significant sub-events within a video class are commonly present among edited videos while less interesting ones appear less frequently, we identify the significant sub-events via a robust recurrent auto-encoder trained on a collection of user-edited videos queried for each particular class of interest. The auto-encoder is trained using a proposed shrinking exponential loss function that makes it robust to noise in the web-crawled training data, and is configured with bidirectional long short term memory (LSTM) LSTM:97 cells to better model the temporal structure of highlight segments. Different from supervised techniques, our method can infer highlights using only a set of downloaded edited videos, without also needing their pre-edited counterparts which are rarely available online. Extensive experiments indicate the promise of our proposed solution in this challenging unsupervised settin",
"We describe a new spatio-temporal video autoencoder, based on a classic spatial image autoencoder and a novel nested temporal autoencoder. The temporal encoder is represented by a differentiable visual memory composed of convolutional long short-term memory (LSTM) cells that integrate changes over time. Here we target motion changes and use as temporal decoder a robust optical flow prediction module together with an image sampler serving as built-in feedback loop. The architecture is end-to-end differentiable. At each time step, the system receives as input a video frame, predicts the optical flow based on the current observation and the LSTM memory state as a dense transformation map, and applies it to the current frame to generate the next frame. By minimising the reconstruction error between the predicted next frame and the corresponding ground truth next frame, we train the whole system to extract features useful for motion estimation without any supervision effort. We present one direct application of the proposed framework in weakly-supervised semantic segmentation of videos through label propagation using optical flow.",
"We use multilayer Long Short Term Memory (LSTM) networks to learn representations of video sequences. Our model uses an encoder LSTM to map an input sequence into a fixed length representation. This representation is decoded using single or multiple decoder LSTMs to perform different tasks, such as reconstructing the input sequence, or predicting the future sequence. We experiment with two kinds of input sequences - patches of image pixels and high-level representations (\"percepts\") of video frames extracted using a pretrained convolutional net. We explore different design choices such as whether the decoder LSTMs should condition on the generated output. We analyze the outputs of the model qualitatively to see how well the model can extrapolate the learned video representation into the future and into the past. We try to visualize and interpret the learned features. We stress test the model by running it on longer time scales and on out-of-domain data. We further evaluate the representations by finetuning them for a supervised learning problem - human action recognition on the UCF-101 and HMDB-51 datasets. We show that the representations help improve classification accuracy, especially when there are only a few training examples. Even models pretrained on unrelated datasets (300 hours of YouTube videos) can help action recognition performance.",
"We use Long Short Term Memory (LSTM) networks to learn representations of video sequences. Our model uses an encoder LSTM to map an input sequence into a fixed length representation. This representation is decoded using single or multiple decoder LSTMs to perform different tasks, such as reconstructing the input sequence, or predicting the future sequence. We experiment with two kinds of input sequences - patches of image pixels and high-level representations (\"percepts\") of video frames extracted using a pretrained convolutional net. We explore different design choices such as whether the decoder LSTMs should condition on the generated output. We analyze the outputs of the model qualitatively to see how well the model can extrapolate the learned video representation into the future and into the past. We further evaluate the representations by finetuning them for a supervised learning problem - human action recognition on the UCF-101 and HMDB-51 datasets. We show that the representations help improve classification accuracy, especially when there are only few training examples. Even models pretrained on unrelated datasets (300 hours of YouTube videos) can help action recognition performance."
]
} |
1901.11168 | 2972498781 | The emergence of continuous health monitoring and the availability of an enormous amount of time series data has provided a great opportunity for the advancement of personal health tracking. In recent years, unsupervised learning methods have drawn special attention of researchers to tackle the sparse annotation of health data and real-time detection of anomalies has been a central problem of interest. However, one problem that has not been well addressed before is the early prediction of forthcoming negative health events. Early signs of an event can introduce subtle and gradual changes in the health signal prior to its onset, detection of which can be invaluable in effective prevention. In this study, we first demonstrate our observations on the shortcoming of widely adopted anomaly detection methods in uncovering the changes prior to a negative health event. We then propose a framework which relies on online clustering of signal segment representations which are automatically learned by a specially designed LSTM auto-encoder. We benchmark our results on the publicly available MIT-PICS dataset and show the effectiveness of our approach by predicting Bradycardia events in infants 1.3 minutes ahead of time with 68 AUC score on average, with no label supervision. Results of our study can indicate the viability of our approach in the early detection of health events in other applications as well. | Prediction of Bradycardia in infants using the PICS dataset was approached before by publishers of the dataset with statistical methods @cite_2 . They specifically used a point process analysis and tried to capture the differences in variance and mean of signal segments before a Bradycardia event. Although this study proves the feasibility and achieves reasonable accuracy, their approach is supervised, hand-engineered, and heavily relies on the observance of multiple onsets of Bradycardia events in each infant, which is not always possible in the real-world setting. This is while our approach focuses on the straightforward collection of normal signals from individuals and the detection of changes in an unsupervised and automatic manner. | {
"cite_N": [
"@cite_2"
],
"mid": [
"2557126235"
],
"abstract": [
"Objective: Episodes of bradycardia are common and recur sporadically in preterm infants, posing a threat to the developing brain and other vital organs. We hypothesize that bradycardias are a result of transient temporal destabilization of the cardiac autonomic control system and that fluctuations in the heart rate signal might contain information that precedes bradycardia. We investigate infant heart rate fluctuations with a novel application of point process theory. Methods: In ten preterm infants, we estimate instantaneous linear measures of the heart rate signal, use these measures to extract statistical features of bradycardia, and propose a simplistic framework for prediction of bradycardia. Results: We present the performance of a prediction algorithm using instantaneous linear measures (mean area under the curve = 0.79 ± 0.018) for over 440 bradycardia events. The algorithm achieves an average forecast time of 116 s prior to bradycardia onset (FPR = 0.15). Our analysis reveals that increased variance in the heart rate signal is a precursor of severe bradycardia. This increase in variance is associated with an increase in power from low content dynamics in the LF band (0.04–0.2 Hz) and lower multiscale entropy values prior to bradycardia. Conclusion: Point process analysis of the heartbeat time series reveals instantaneous measures that can be used to predict infant bradycardia prior to onset. Significance: Our findings are relevant to risk stratification, predictive monitoring, and implementation of preventative strategies for reducing morbidity and mortality associated with bradycardia in neonatal intensive care units."
]
} |
1901.11220 | 2912240042 | Initial access (IA) is a fundamental procedure in cellular systems where user equipment (UE) detects base station (BS) and acquires synchronization. Due to the necessity of using antenna arrays for IA in millimeter-wave (mmW) systems, BS simultaneously performs beam training to acquire angular channel state information. The state-of-the-art directional IA (DIA) uses a set of narrow sounding beams in IA, where different beam pairs are sequentially measured, and the best candidate is determined. However, the directional beam training accuracy depends on scanning beam angular resolution, and consequently its improvement requires additional dedicated radio resources, access latency, and overhead. To remedy the problem of access latency and overhead in DIA, this paper proposes to use quasi-omni pseudorandom sounding beams for IA, and develops an algorithm for joint initial access and fine resolution initial beam training without requiring additional radio resources. It comprehensively models realistic timing and frequency synchronization errors encountered in IA. We provide the analysis of the proposed algorithm's miss detection rate under timing synchronization errors, and we further derive Cramer–Rao lower bound of angular estimation under frequency offset, considering the 5G-NR compliant IA procedure. To accommodate the ever increasing bandwidth for beam training in standard evolution beyond 5G, we design the beam squint robust algorithm. For realistic performance evaluation under mmW channels, we use QuaDRiGa simulator with mmMAGIC model at 28 GHz to show that the proposed approach is advantageous to DIA. The proposed algorithm offers orders of magnitude access latency saving compared to DIA, when the same discovery, post training SNR, and overhead performance are targeted. This conclusion holds true in various propagation environments and three-dimensional locations of a mmW pico-cell with up to 140 m radius. Furthermore, our results demonstrate that the proposed beam squint robust algorithm is able to retain unaffected performance with increased beam training bandwidth. | The alternative approaches for beam training are based on parametric channel estimation @cite_31 @cite_24 @cite_32 @cite_18 @cite_25 @cite_17 @cite_42 @cite_45 . Exploiting the mmW sparse scattering nature, compressive sensing (CS) approaches have been considered to effectively estimate channel parameters based on channel observations obtained via various sounding beams. Works @cite_31 @cite_24 proposed a CS-based narrowband BF training with pseudorandom sounding beamformers in the downlink, and @cite_32 extended this approach for a wideband channel. Other related works include channel covariance estimation @cite_18 @cite_25 @cite_17 which requires periodic channel observations, and UE centric uplink training @cite_42 @cite_45 . It is worth nothing that all recent works focus on channel estimation alone while assuming perfect cell discovery and synchronization. The 5G-NR frame structure that supports IA is rarely considered, and further the feasibility of joint initial access and CS-based beam training has not been investigated. | {
"cite_N": [
"@cite_18",
"@cite_42",
"@cite_32",
"@cite_24",
"@cite_45",
"@cite_31",
"@cite_25",
"@cite_17"
],
"mid": [
"2015301486",
"624827785",
"2964311703",
"2339667469",
"2784857961",
"2611243941",
"2288934705",
"2795172492"
],
"abstract": [
"We propose joint spatial division and multiplexing (JSDM), an approach to multiuser MIMO downlink that exploits the structure of the correlation of the channel vectors in order to allow for a large number of antennas at the base station while requiring reduced-dimensional channel state information at the transmitter (CSIT). JSDM achieves significant savings both in the downlink training and in the CSIT uplink feedback, thus making the use of large antenna arrays at the base station potentially suitable also for frequency division duplexing (FDD) systems, for which uplink downlink channel reciprocity cannot be exploited. In the proposed scheme, the multiuser MIMO downlink precoder is obtained by concatenating a prebeamforming matrix, which depends only on the channel second-order statistics, with a classical multiuser precoder, based on the instantaneous knowledge of the resulting reduced dimensional “effective” channel matrix. We prove a simple condition under which JSDM incurs no loss of optimality with respect to the full CSIT case. For linear uniformly spaced arrays, we show that such condition is approached in the large number of antennas limit. For this case, we use Szego's asymptotic theory of Toeplitz matrices to show that a DFT-based prebeamforming matrix is near-optimal, requiring only coarse information about the users angles of arrival and angular spread. Finally, we extend these ideas to the case of a 2-D base station antenna array, with 3-D beamforming, including multiple beams in the elevation angle direction. We provide guidelines for the prebeamforming optimization and calculate the system spectral efficiency under proportional fairness and max-min fairness criteria, showing extremely attractive performance. Our numerical results are obtained via asymptotic random matrix theory, avoiding lengthy Monte Carlo simulations and providing accurate results for realistic (finite) number of antennas and users.",
"We propose and investigate a compressive architecture for estimation and tracking of sparse spatial channels in millimeter (mm) wave picocellular networks. The base stations are equipped with antenna arrays with a large number of elements (which can fit within compact form factors because of the small carrier wavelength) and employ radio frequency (RF) beamforming, so that standard least squares adaptation techniques (which require access to individual antenna elements) are not applicable. We focus on the downlink, and show that “compressive beacons,” transmitted using pseudorandom phase settings at the base station array, and compressively processed using pseudorandom phase settings at the mobile array, provide information sufficient for accurate estimation of the two-dimensional (2D) spatial frequencies associated with the directions of departure of the dominant rays from the base station, and the associated complex gains. This compressive approach is compatible with coarse phase-only control, and is based on a near-optimal sequential algorithm for frequency estimation which approaches the Cramer Rao Lower Bound. The algorithm exploits the geometric continuity of the channel across successive beaconing intervals to reduce the overhead to less than 1 even for very large ( @math ) arrays. Compressive beaconing is essentially omnidirectional, and hence does not enjoy the SNR and spatial reuse benefits of beamforming obtained during data transmission. We therefore discuss system level design considerations for ensuring that the beacon SNR is sufficient for accurate channel estimation, and that inter-cell beacon interference is controlled by an appropriate reuse scheme.",
"Millimeter-wave (mm-wave) frequency bands provide an opportunity for much wider channel bandwidth compared with the traditional sub-6-GHz band. Communication at mm-waves is, however, quite challenging due to the severe propagation pathloss incurred by conventional isotropic antennas. To cope with this problem, directional beamforming both at the base station (BS) side and at the user equipment (UE) side is necessary in order to establish a strong path conveying enough signal power. Finding such beamforming directions is referred to as beam alignment (BA). This paper presents a new scheme for efficient BA. Our scheme finds a strong propagation path identified by an angle-of-arrival (AoA) and angle-of-departure (AoD) pair, by exploring the AoA–AoD domain through pseudo-random multi-finger beam patterns and constructing an estimate of the resulting second-order statistics (namely, the average received power for each pseudo-random beam configuration). The resulting under-determined system of equations is efficiently solved using non-negative constrained least-squares, yielding naturally a sparse non-negative vector solution whose maximum component identifies the optimal path. As a result, our scheme is highly robust to variations of the channel time dynamics compared with alternative concurrent approaches based on the estimation of the instantaneous channel coefficients, rather than of their second-order statistics. In the proposed scheme, the BS probes the channel in the downlink and trains simultaneously an arbitrarily large number of UEs. Thus, “beam refinement,” with multiple interactive rounds of downlink uplink transmissions, is not needed. This results in a scalable BA protocol, where the protocol overhead is virtually independent of the number of UEs, since all the UEs run the BA procedure at the same time. Extensive simulation results illustrate that our approach is superior to the state-of-the-art BA schemes proposed in the literature in terms of training overhead in multi-user scenarios and robustness to variations in the channel dynamics.",
"We propose an efficient open-loop channel estimator for a millimeter-wave (mm-wave) hybrid multiple-input multiple-output (MIMO) system consisting of radio-frequency (RF) beamformers with large antenna arrays followed by a baseband MIMO processor. A sparse signal recovery problem exploiting the sparse nature of mm-wave channels is formulated for channel estimation based on the parametric channel model with quantized angles of departures arrivals (AoDs AoAs), called the angle grids. The problem is solved by the orthogonal matching pursuit (OMP) algorithm employing a redundant dictionary consisting of array response vectors with finely quantized angle grids. We suggest the use of non-uniformly quantized angle grids and show that such grids reduce the coherence of the redundant dictionary. The lower and upper bounds of the sum-of-squared errors of the proposed OMP-based estimator are derived analytically: the lower bound is derived by considering the oracle estimator that assumes the knowledge of AoDs AoAs, and the upper bound is derived based on the results of the OMP performance guarantees. The design of training vectors (or sensing matrix) is particularly important in hybrid MIMO systems, because the RF beamformer prevents the use of independent and identically distributed random training vectors, which are popular in compressed sensing. We design training vectors so that the total coherence of the equivalent sensing matrix is minimized for a given RF beamforming matrix, which is assumed to be unitary. It is observed that the estimation accuracy can be improved significantly by randomly permuting the columns of the RF beamforming matrix. The simulation results demonstrate the advantage of the proposed OMP with a redundant dictionary over the existing methods such as the least squares method and the OMP based on the virtual channel model.",
"Millimeter (mm) wave massive MIMO has the potential for delivering orders of magnitude increases in mobile data rates, with compact antenna arrays providing narrow steerable beams for unprecedented levels of spatial reuse. A fundamental technical bottleneck, however, is rapid spatial channel estimation and beam adaptation in the face of mobility and blockage. Recently proposed compressive techniques which exploit the sparsity of mm wave channels are a promising approach to this problem, with overhead scaling linearly with the number of dominant paths and logarithmically with the number of array elements. Further, they can be implemented with RF beamforming with low-precision phase control. However, these methods make implicit assumptions on long-term phase coherence that are not satisfied by existing hardware. In this paper, we propose and evaluate a noncoherent compressive channel estimation technique which can estimate a sparse spatial channel based on received signal strength (RSS) alone, and is compatible with off-the-shelf hardware. The approach is based on cascading phase retrieval (i.e., recovery of complex-valued measurements from RSS measurements, up to a scalar multiple) with coherent compressive estimation. While a conventional cascade scheme would multiply two measurement matrices to obtain an overall matrix whose entries are in a continuum, a key novelty in our scheme is that we constrain the overall measurement matrix to be implementable using coarsely quantized pseudorandom phases, employing a virtual decomposition of the matrix into a product of measurement matrices for phase retrieval and compressive estimation. Theoretical and simulation results show that our noncoherent method scales almost as well with array size as its coherent counterpart, thus inheriting the scalability and low overhead of the latter.",
"The abundant spectrum at millimeter-wave (mmWave) has the potential to greatly increase the capacity of 5G cellular systems. However, to overcome the high pathloss in the mmWave frequencies, beamforming with large antenna arrays is required at both the base station and user equipments for sufficient link budget. This feature is a challenge for beamforming training during initial access due to low SNR and poor synchronization. A recently developed compressive sensing (CS) based training algorithm exploits channel sparsity but it is vulnerable to phase error from poor synchronization. We propose a novel CS-based algorithm that tracks and compensates frequency offset and phase noise. Simulation results show that the proposed method improves achievable rate by 10 times compared with existing CS-based method during initial beamforming training.",
"We consider distributed transmit beamforming from a cluster of cooperating nodes towards a distant destination, over a wideband dispersive channel. We consider explicit aggregate feedback, with the destination broadcasting its feedback to all nodes in the transmit cluster. Explicit feedback allows for frequency division duplex (FDD) operation, since there is no reliance on channel reciprocity. Aggregate feedback enables scalability, since the destination is agnostic to the number and identity of the transmitters. There are two key contributions. First, it is shown in a narrowband setting that, unlike a well-known one bit aggregate feedback scheme, a training-based approach in which each transmitter learns its channel to the destination based on aggregate feedback, is resilient to noise, even when the feedback is heavily quantized. Second, extending the scheme to wideband settings by using OFDM, training on a subset of subcarriers, we show that interpolation across subcarriers using sparse time domain modeling provides accurate channel estimates. In particular, at low SNR per node, our approach, which exploits the correlation across subcarriers, outperforms independent per- subcarrier channel estimates in a hypothetical benchmark system in which training and feedback is applied to all subcarriers. Our simulation results show that the beamforming performance obtained by our approach is a substantial fraction of the ideal beamforming gain, even for typical received SNRs per node as low as −20 dB, and coarse 4-phase feedback quantization. In general, the method requires training time linearly proportional to the number of cooperating nodes and inversely proportional to the per-node SNR.",
"In millimeter wave (mm-wave) massive multiple-input multiple-output (MIMO) systems, acquiring accurate channel state information is essential for efficient beamforming (BF) and multiuser interference cancellation, which is a challenging task since a low signal-to-noise ratio is encountered before BF in large antenna arrays. The mm-wave channel exhibits a 3-D clustered structure in the virtual angle of arrival (AOA), angle of departure (AOD), and delay domain that is imposed by the effect of power leakage, angular spread, and cluster duration. We extend the approximate message passing (AMP) with a nearest neighbor pattern learning algorithm for improving the attainable channel estimation performance, which adaptively learns and exploits the clustered structure in the 3-D virtual AOA-AOD-delay domain. The proposed method is capable of approaching the performance bound described by the state evolution based on vector AMP framework, and our simulation results verify its superiority in mm-wave systems associated with a broad bandwidth."
]
} |
1901.11220 | 2912240042 | Initial access (IA) is a fundamental procedure in cellular systems where user equipment (UE) detects base station (BS) and acquires synchronization. Due to the necessity of using antenna arrays for IA in millimeter-wave (mmW) systems, BS simultaneously performs beam training to acquire angular channel state information. The state-of-the-art directional IA (DIA) uses a set of narrow sounding beams in IA, where different beam pairs are sequentially measured, and the best candidate is determined. However, the directional beam training accuracy depends on scanning beam angular resolution, and consequently its improvement requires additional dedicated radio resources, access latency, and overhead. To remedy the problem of access latency and overhead in DIA, this paper proposes to use quasi-omni pseudorandom sounding beams for IA, and develops an algorithm for joint initial access and fine resolution initial beam training without requiring additional radio resources. It comprehensively models realistic timing and frequency synchronization errors encountered in IA. We provide the analysis of the proposed algorithm's miss detection rate under timing synchronization errors, and we further derive Cramer–Rao lower bound of angular estimation under frequency offset, considering the 5G-NR compliant IA procedure. To accommodate the ever increasing bandwidth for beam training in standard evolution beyond 5G, we design the beam squint robust algorithm. For realistic performance evaluation under mmW channels, we use QuaDRiGa simulator with mmMAGIC model at 28 GHz to show that the proposed approach is advantageous to DIA. The proposed algorithm offers orders of magnitude access latency saving compared to DIA, when the same discovery, post training SNR, and overhead performance are targeted. This conclusion holds true in various propagation environments and three-dimensional locations of a mmW pico-cell with up to 140 m radius. Furthermore, our results demonstrate that the proposed beam squint robust algorithm is able to retain unaffected performance with increased beam training bandwidth. | There are also recent works that consider some practical aspects of IA. For example, frequency offset robust algorithms in narrowband mmW beam training are reported in @cite_6 @cite_40 @cite_29 . There are several hardware prototypes that consider a practical approach of using received signal strength (RSS) in CS-based beam training. Channel estimation problem without phase measurement is a challenging problem, which was solved via novel signal processing algorithms based on RSS matching pursuit @cite_10 , Hash table @cite_5 , and sparse phase retrieval @cite_28 . Note that phase free measurements were associated with a particular testbed, and this constraint does not necessarily apply to mmW systems in general. In summary, while IA and beam training algorithms have been extensively studied in the literature, there is a lack of understanding about the theoretical limits and signal processing algorithms that jointly achieve cell discovery and accurate BF training using asynchronous IA signal in mmW frequency selective channel. | {
"cite_N": [
"@cite_28",
"@cite_29",
"@cite_6",
"@cite_40",
"@cite_5",
"@cite_10"
],
"mid": [
"2784857961",
"2795172492",
"2964311703",
"2611243941",
"2075328041",
"2964056649"
],
"abstract": [
"Millimeter (mm) wave massive MIMO has the potential for delivering orders of magnitude increases in mobile data rates, with compact antenna arrays providing narrow steerable beams for unprecedented levels of spatial reuse. A fundamental technical bottleneck, however, is rapid spatial channel estimation and beam adaptation in the face of mobility and blockage. Recently proposed compressive techniques which exploit the sparsity of mm wave channels are a promising approach to this problem, with overhead scaling linearly with the number of dominant paths and logarithmically with the number of array elements. Further, they can be implemented with RF beamforming with low-precision phase control. However, these methods make implicit assumptions on long-term phase coherence that are not satisfied by existing hardware. In this paper, we propose and evaluate a noncoherent compressive channel estimation technique which can estimate a sparse spatial channel based on received signal strength (RSS) alone, and is compatible with off-the-shelf hardware. The approach is based on cascading phase retrieval (i.e., recovery of complex-valued measurements from RSS measurements, up to a scalar multiple) with coherent compressive estimation. While a conventional cascade scheme would multiply two measurement matrices to obtain an overall matrix whose entries are in a continuum, a key novelty in our scheme is that we constrain the overall measurement matrix to be implementable using coarsely quantized pseudorandom phases, employing a virtual decomposition of the matrix into a product of measurement matrices for phase retrieval and compressive estimation. Theoretical and simulation results show that our noncoherent method scales almost as well with array size as its coherent counterpart, thus inheriting the scalability and low overhead of the latter.",
"In millimeter wave (mm-wave) massive multiple-input multiple-output (MIMO) systems, acquiring accurate channel state information is essential for efficient beamforming (BF) and multiuser interference cancellation, which is a challenging task since a low signal-to-noise ratio is encountered before BF in large antenna arrays. The mm-wave channel exhibits a 3-D clustered structure in the virtual angle of arrival (AOA), angle of departure (AOD), and delay domain that is imposed by the effect of power leakage, angular spread, and cluster duration. We extend the approximate message passing (AMP) with a nearest neighbor pattern learning algorithm for improving the attainable channel estimation performance, which adaptively learns and exploits the clustered structure in the 3-D virtual AOA-AOD-delay domain. The proposed method is capable of approaching the performance bound described by the state evolution based on vector AMP framework, and our simulation results verify its superiority in mm-wave systems associated with a broad bandwidth.",
"Millimeter-wave (mm-wave) frequency bands provide an opportunity for much wider channel bandwidth compared with the traditional sub-6-GHz band. Communication at mm-waves is, however, quite challenging due to the severe propagation pathloss incurred by conventional isotropic antennas. To cope with this problem, directional beamforming both at the base station (BS) side and at the user equipment (UE) side is necessary in order to establish a strong path conveying enough signal power. Finding such beamforming directions is referred to as beam alignment (BA). This paper presents a new scheme for efficient BA. Our scheme finds a strong propagation path identified by an angle-of-arrival (AoA) and angle-of-departure (AoD) pair, by exploring the AoA–AoD domain through pseudo-random multi-finger beam patterns and constructing an estimate of the resulting second-order statistics (namely, the average received power for each pseudo-random beam configuration). The resulting under-determined system of equations is efficiently solved using non-negative constrained least-squares, yielding naturally a sparse non-negative vector solution whose maximum component identifies the optimal path. As a result, our scheme is highly robust to variations of the channel time dynamics compared with alternative concurrent approaches based on the estimation of the instantaneous channel coefficients, rather than of their second-order statistics. In the proposed scheme, the BS probes the channel in the downlink and trains simultaneously an arbitrarily large number of UEs. Thus, “beam refinement,” with multiple interactive rounds of downlink uplink transmissions, is not needed. This results in a scalable BA protocol, where the protocol overhead is virtually independent of the number of UEs, since all the UEs run the BA procedure at the same time. Extensive simulation results illustrate that our approach is superior to the state-of-the-art BA schemes proposed in the literature in terms of training overhead in multi-user scenarios and robustness to variations in the channel dynamics.",
"The abundant spectrum at millimeter-wave (mmWave) has the potential to greatly increase the capacity of 5G cellular systems. However, to overcome the high pathloss in the mmWave frequencies, beamforming with large antenna arrays is required at both the base station and user equipments for sufficient link budget. This feature is a challenge for beamforming training during initial access due to low SNR and poor synchronization. A recently developed compressive sensing (CS) based training algorithm exploits channel sparsity but it is vulnerable to phase error from poor synchronization. We propose a novel CS-based algorithm that tracks and compensates frequency offset and phase noise. Simulation results show that the proposed method improves achievable rate by 10 times compared with existing CS-based method during initial beamforming training.",
"In this paper, we propose jointly optimized linear transceiver algorithms called interference aware-coordinated beamforming (IA-CBF) for a two-cell system where each base station is equipped with multiple transmit antennas. To generalize IA-CBF to more than two-cell scenarios, a new beam-switching mechanism combined with IA-CBF is proposed. For a two-cell system, we derive a minimum-mean-square-error-type IA-CBF algorithm based on a lower bound on the achievable sum rate. We propose optimal (under an assumption of zero other-cell interference) and suboptimal transmit receive beamforming vectors through zero-forcing IA-CBF algorithms. We also investigate the optimality of the proposed IA-CBF algorithms with respect to the number of receive antennas. Numerical results confirm that the proposed system with two transmit receive antennas achieves the full degrees of freedom (a.k.a. multiplexing gain) of the two-cell multiple-input multiple-output channel while showing a better sum rate performance than competitive solutions such as non-cooperative eigen-beamforming and interference nulling. A three-dimensional ray tracing tool is also used to evaluate the proposed multi-cell IA-CBF algorithm.",
"We study RF-enabled wireless energy transfer (WET) via energy beamforming, from a multi-antenna energy transmitter (ET) to multiple energy receivers (ERs) in a backscatter communication system such as RFID. The acquisition of the forward-channel (i.e., ET-to-ER) state information (F-CSI) at the ET (or RFID reader) is challenging, since the ERs (or RFID tags) are typically too energy-and-hardware-constrained to estimate or feedback the F-CSI. The ET leverages its observed backscatter signals to estimate the backscatter-channel (i.e., ET-to-ER-to-ET) state information (BS-CSI) directly. We first analyze the harvested energy obtained using the estimated BS-CSI. Furthermore, we optimize the resource allocation to maximize the total utility of harvested energy. For WET to single ER, we obtain the optimal channel-training energy in a semiclosed form. For WET to multiple ERs, we optimize the channel-training energy and the energy allocation weights for different energy beams. For the straightforward weighted-sum-energy (WSE) maximization, the optimal WET scheme is shown to use only one energy beam, which leads to unfairness among ERs and motivates us to consider the complicated proportional-fair-energy (PFE) maximization. For PFE maximization, we show that it is a biconvex problem, and propose a block-coordinate-descent-based algorithm to find the close-to-optimal solution. Numerical results show that with the optimized solutions, the harvested energy suffers slight reduction of less than 10 , compared to that obtained using the perfect F-CSI."
]
} |
1901.11188 | 2952085755 | The vulnerability of neural networks under adversarial attacks has raised serious concerns and motivated extensive research. It has been shown that both neural networks and adversarial attacks against them can be sensitive to input transformations such as linear translation and rotation, and that human vision, which is robust against adversarial attacks, is invariant to natural input transformations. Based on these, this paper tests the hypothesis that model robustness can be further improved when it is adversarially trained against transformed attacks and transformation-invariant attacks. Experiments on MNIST, CIFAR-10, and restricted ImageNet show that while transformations of attacks alone do not affect robustness, transformation-invariant attacks can improve model robustness by 2.5 on MNIST, 3.7 on CIFAR-10, and 1.1 on restricted ImageNet. We discuss the intuition behind this phenomenon. | @cite_25 proposed using random transformations to pre-processing the input images to improve model robustness. It was later shown, however, that this approach creates a gradient masking effect and can be broken by robust attacks @cite_42 . Unlike @cite_25 , we consider the transformation as part of our model during the adversarial training process. | {
"cite_N": [
"@cite_42",
"@cite_25"
],
"mid": [
"2795995837",
"2964077693"
],
"abstract": [
"Deep neural networks are known to be vulnerable to adversarial examples, i.e., images that are maliciously perturbed to fool the model. Generating adversarial examples has been mostly limited to finding small perturbations that maximize the model prediction error. Such images, however, contain artificial perturbations that make them somewhat distinguishable from natural images. This property is used by several defense methods to counter adversarial examples by applying denoising filters or training the model to be robust to small perturbations. In this paper, we introduce a new class of adversarial examples, namely \"Semantic Adversarial Examples,\" as images that are arbitrarily perturbed to fool the model, but in such a way that the modified image semantically represents the same object as the original image. We formulate the problem of generating such images as a constrained optimization problem and develop an adversarial transformation based on the shape bias property of human cognitive system. In our method, we generate adversarial images by first converting the RGB image into the HSV (Hue, Saturation and Value) color space and then randomly shifting the Hue and Saturation components, while keeping the Value component the same. Our experimental results on CIFAR10 dataset show that the accuracy of VGG16 network on adversarial color-shifted images is 5.7 .",
"Deep neural networks are known to be vulnerable to adversarial examples, i.e., images that are maliciously perturbed to fool the model. Generating adversarial examples has been mostly limited to finding small perturbations that maximize the model prediction error. Such images, however, contain artificial perturbations that make them somewhat distinguishable from natural images. This property is used by several defense methods to counter adversarial examples by applying denoising filters or training the model to be robust to small perturbations. In this paper, we introduce a new class of adversarial examples, namely \"Semantic Adversarial Examples,\" as images that are arbitrarily perturbed to fool the model, but in such a way that the modified image semantically represents the same object as the original image. We formulate the problem of generating such images as a constrained optimization problem and develop an adversarial transformation based on the shape bias property of human cognitive system. In our method, we generate adversarial images by first converting the RGB image into the HSV (Hue, Saturation and Value) color space and then randomly shifting the Hue and Saturation components, while keeping the Value component the same. Our experimental results on CIFAR10 dataset show that the accuracy of VGG16 network on adversarial color-shifted images is 5.7 ."
]
} |
1901.11188 | 2952085755 | The vulnerability of neural networks under adversarial attacks has raised serious concerns and motivated extensive research. It has been shown that both neural networks and adversarial attacks against them can be sensitive to input transformations such as linear translation and rotation, and that human vision, which is robust against adversarial attacks, is invariant to natural input transformations. Based on these, this paper tests the hypothesis that model robustness can be further improved when it is adversarially trained against transformed attacks and transformation-invariant attacks. Experiments on MNIST, CIFAR-10, and restricted ImageNet show that while transformations of attacks alone do not affect robustness, transformation-invariant attacks can improve model robustness by 2.5 on MNIST, 3.7 on CIFAR-10, and 1.1 on restricted ImageNet. We discuss the intuition behind this phenomenon. | Attacks from an ensemble of black-box models have been used to effectively avoid gradient masking in one-step adversarial training @cite_28 . While our model also uses an ensemble of attacks, these attacks are white-box and multi-step. Importantly, these attacks do not cause gradient masking. | {
"cite_N": [
"@cite_28"
],
"mid": [
"2949358371"
],
"abstract": [
"Solving for adversarial examples with projected gradient descent has been demonstrated to be highly effective in fooling the neural network based classifiers. However, in the black-box setting, the attacker is limited only to the query access to the network and solving for a successful adversarial example becomes much more difficult. To this end, recent methods aim at estimating the true gradient signal based on the input queries but at the cost of excessive queries. We propose an efficient discrete surrogate to the optimization problem which does not require estimating the gradient and consequently becomes free of the first order update hyperparameters to tune. Our experiments on Cifar-10 and ImageNet show the state of the art black-box attack performance with significant reduction in the required queries compared to a number of recently proposed methods. The source code is available at this https URL."
]
} |
1901.11167 | 2952180917 | In the literature, tensors have been effectively used for capturing the context information in language models. However, the existing methods usually adopt relatively-low order tensors, which have limited expressive power in modeling language. Developing a higher-order tensor representation is challenging, in terms of deriving an effective solution and showing its generality. In this paper, we propose a language model named Tensor Space Language Model (TSLM), by utilizing tensor networks and tensor decomposition. In TSLM, we build a high-dimensional semantic space constructed by the tensor product of word vectors. Theoretically, we prove that such tensor representation is a generalization of the n-gram language model. We further show that this high-order tensor representation can be decomposed to a recursive calculation of conditional probability for language modeling. The experimental results on Penn Tree Bank (PTB) dataset and WikiText benchmark demonstrate the effectiveness of TSLM. | There have been tremendous research efforts in the field of statistical language modeling. Some earlier language models are based on the Markov assumption are represented by @math -gram models @cite_17 , where the prediction of the next word is often conditioned just on @math preceding words. For @math -gram models, Kneser and Ney proposed the most well-known KN smoothing method, and some researchers continued to improve the smoothing method, as well as introduced the low-rank model. Neural Probabilistic Language Model @cite_18 is to learn the joint probability function of sequence of words in a language, which shows the improvement on @math -gram models. Recently, RNN @cite_10 and Long Short-Term Memory (LSTM) networks @cite_1 achieve promising results on language model tasks. | {
"cite_N": [
"@cite_1",
"@cite_18",
"@cite_10",
"@cite_17"
],
"mid": [
"2158195707",
"2091812280",
"2120861206",
"2950797609"
],
"abstract": [
"We survey the most widely-used algorithms for smoothing models for language n -gram modeling. We then present an extensive empirical comparison of several of these smoothing techniques, including those described by Jelinek and Mercer (1980); Katz (1987); Bell, Cleary and Witten (1990); Ney, Essen and Kneser (1994), and Kneser and Ney (1995). We investigate how factors such as training data size, training corpus (e.g. Brown vs. Wall Street Journal), count cutoffs, and n -gram order (bigram vs. trigram) affect the relative performance of these methods, which is measured through the cross-entropy of test data. We find that these factors can significantly affect the relative performance of models, with the most significant factor being training data size. Since no previous comparisons have examined these factors systematically, this is the first thorough characterization of the relative performance of various algorithms. In addition, we introduce methodologies for analyzing smoothing algorithm efficacy in detail, and using these techniques we motivate a novel variation of Kneser?Ney smoothing that consistently outperforms all other algorithms evaluated. Finally, results showing that improved language model smoothing leads to improved speech recognition performance are presented.",
"The supremacy of n-gram models in statistical language modelling has recently been challenged by parametric models that use distributed representations to counteract the difficulties caused by data sparsity. We propose three new probabilistic language models that define the distribution of the next word in a sequence given several preceding words by using distributed representations of those words. We show how real-valued distributed representations for words can be learned at the same time as learning a large set of stochastic binary hidden features that are used to predict the distributed representation of the next word from previous distributed representations. Adding connections from the previous states of the binary hidden features improves performance as does adding direct connections between the real-valued distributed representations. One of our models significantly outperforms the very best n-gram models.",
"In spite of their superior performance, neural probabilistic language models (NPLMs) remain far less widely used than n-gram models due to their notoriously long training times, which are measured in weeks even for moderately-sized datasets. Training NPLMs is computationally expensive because they are explicitly normalized, which leads to having to consider all words in the vocabulary when computing the log-likelihood gradients. We propose a fast and simple algorithm for training NPLMs based on noise-contrastive estimation, a newly introduced procedure for estimating unnormalized continuous distributions. We investigate the behaviour of the algorithm on the Penn Treebank corpus and show that it reduces the training times by more than an order of magnitude without affecting the quality of the resulting models. The algorithm is also more efficient and much more stable than importance sampling because it requires far fewer noise samples to perform well. We demonstrate the scalability of the proposed approach by training several neural language models on a 47M-word corpus with a 80K-word vocabulary, obtaining state-of-the-art results on the Microsoft Research Sentence Completion Challenge dataset.",
"In spite of their superior performance, neural probabilistic language models (NPLMs) remain far less widely used than n-gram models due to their notoriously long training times, which are measured in weeks even for moderately-sized datasets. Training NPLMs is computationally expensive because they are explicitly normalized, which leads to having to consider all words in the vocabulary when computing the log-likelihood gradients. We propose a fast and simple algorithm for training NPLMs based on noise-contrastive estimation, a newly introduced procedure for estimating unnormalized continuous distributions. We investigate the behaviour of the algorithm on the Penn Treebank corpus and show that it reduces the training times by more than an order of magnitude without affecting the quality of the resulting models. The algorithm is also more efficient and much more stable than importance sampling because it requires far fewer noise samples to perform well. We demonstrate the scalability of the proposed approach by training several neural language models on a 47M-word corpus with a 80K-word vocabulary, obtaining state-of-the-art results on the Microsoft Research Sentence Completion Challenge dataset."
]
} |
1901.10951 | 2914704920 | For autonomous vehicles to be able to operate successfully they need to be aware of other vehicles with sufficient time to make safe, stable plans. Given the possible closing speeds between two vehicles, this necessitates the ability to accurately detect distant vehicles. Many current image-based object detectors using convolutional neural networks exhibit excellent performance on existing datasets such as KITTI. However, the performance of these networks falls when detecting small (distant) objects. We demonstrate that incorporating radar data can boost performance in these difficult situations. We also introduce an efficient automated method for training data generation using cameras of different focal lengths. | Object detection has been a major topic of computer vision research and over recent years a number of fully convolutional object detectors have been proposed. Two-stage methods, in particular Faster R-CNN @cite_2 , provide state-of-the-art performance but are computationally expensive. One-stage methods @cite_17 @cite_3 are structurally simpler and can to operate in real-time but suffer a performance penalty. To improve performance on difficult examples and bridge the performance gap between two-stage and one-stage detectors, Lin @cite_5 propose a loss function that focuses the loss on examples about which the classifier is least confident. This however relies on the labels being highly accurate which may not be the case if they are automatically generated. | {
"cite_N": [
"@cite_5",
"@cite_17",
"@cite_3",
"@cite_2"
],
"mid": [
"2949232997",
"2592384071",
"2589615404",
"2963721253"
],
"abstract": [
"Detection of partially occluded objects is a challenging computer vision problem. Standard Convolutional Neural Network (CNN) detectors fail if parts of the detection window are occluded, since not every sub-part of the window is discriminative on its own. To address this issue, we propose a novel loss layer for CNNs, named grid loss, which minimizes the error rate on sub-blocks of a convolution layer independently rather than over the whole feature map. This results in parts being more discriminative on their own, enabling the detector to recover if the detection window is partially occluded. By mapping our loss layer back to a regular fully connected layer, no additional computational cost is incurred at runtime compared to standard CNNs. We demonstrate our method for face detection on several public face detection benchmarks and show that our method outperforms regular CNNs, is suitable for realtime applications and achieves state-of-the-art performance.",
"Object detection when provided image-level labels instead of instance-level labels (i.e., bounding boxes) during training is an important problem in computer vision, since large scale image datasets with instance-level labels are extremely costly to obtain. In this paper, we address this challenging problem by developing an Expectation-Maximization (EM) based object detection method using deep convolutional neural networks (CNNs). Our method is applicable to both the weakly-supervised and semi-supervised settings. Extensive experiments on PASCAL VOC 2007 benchmark show that (1) in the weakly supervised setting, our method provides significant detection performance improvement over current state-of-the-art methods, (2) having access to a small number of strongly (instance-level) annotated images, our method can almost match the performace of the fully supervised Fast RCNN. We share our source code at this https URL",
"Object proposals have recently emerged as an essential cornerstone for object detection. The current state-of-the-art object detectors employ object proposals to detect objects within a modest set of candidate bounding box proposals instead of exhaustively searching across an image using the sliding window approach. However, achieving high recall and good localization with few proposals is still a challenging problem. The challenge becomes even more difficult in the context of autonomous driving, in which small objects, occlusion, shadows, and reflections usually occur. In this paper, we present a robust object proposals re-ranking algorithm that effectivity re-ranks candidates generated from a customized class-independent 3DOP (3D Object Proposals) method using a two-stream convolutional neural network (CNN). The goal is to ensure that those proposals that accurately cover the desired objects are amongst the few top-ranked candidates. The proposed algorithm, which we call DeepStereoOP, exploits not only RGB images as in the conventional CNN architecture, but also depth features including disparity map and distance to the ground. Experiments show that the proposed algorithm outperforms all existing object proposal algorithms on the challenging KITTI benchmark in terms of both recall and localization. Furthermore, the combination of DeepStereoOP and Fast R-CNN achieves one of the best detection results of all three KITTI object classes. HighlightsWe present a robust object proposals re-ranking algorithm for object detection in autonomous driving.Both RGB images and depth features are included in the proposed two-stream CNN architecture called DeepStereoOP.Initial object proposals are generated from a customized class-independent 3DOP method.Experiments show that the proposed algorithm outperforms all existing object proposals algorithms.The combination of DeepStereoOP and Fast R-CNN achieves one of the best detection results on KITTI benchmark.",
"This paper proposes a computationally efficient approach to detecting objects natively in 3D point clouds using convolutional neural networks (CNNs). In particular, this is achieved by leveraging a feature-centric voting scheme to implement novel convolutional layers which explicitly exploit the sparsity encountered in the input. To this end, we examine the trade-off between accuracy and speed for different architectures and additionally propose to use an L 1 penalty on the filter activations to further encourage sparsity in the intermediate representations. To the best of our knowledge, this is the first work to propose sparse convolutional layers and L 1 regularisation for efficient large-scale processing of 3D data. We demonstrate the efficacy of our approach on the KITTI object detection benchmark and show that VoteSDeep models with as few as three layers outperform the previous state of the art in both laser and laser-vision based approaches by margins of up to 40 while remaining highly competitive in terms of processing time."
]
} |
1901.10951 | 2914704920 | For autonomous vehicles to be able to operate successfully they need to be aware of other vehicles with sufficient time to make safe, stable plans. Given the possible closing speeds between two vehicles, this necessitates the ability to accurately detect distant vehicles. Many current image-based object detectors using convolutional neural networks exhibit excellent performance on existing datasets such as KITTI. However, the performance of these networks falls when detecting small (distant) objects. We demonstrate that incorporating radar data can boost performance in these difficult situations. We also introduce an efficient automated method for training data generation using cameras of different focal lengths. | A common approach when training object detectors for a specific task is to pre-train a feature extractor using ImageNet @cite_18 and then fine tune the features with the limited training data available for the task. In @cite_10 Shen show that, given careful network design, it is possible to obtain state of the art results without this pre-training process. This implies that a fusion network, such as the one we are proposing, is not at an insurmountable disadvantage if pre-training is not performed. | {
"cite_N": [
"@cite_18",
"@cite_10"
],
"mid": [
"2750432752",
"2777511827"
],
"abstract": [
"Current CNN based object detectors need initialization from pre-trained ImageNet classification models, which are usually time-consuming. In this paper, we present a fully convolutional feature mimic framework to train very efficient CNN based detectors, which do not need ImageNet pre-training and achieve competitive performance as the large and slow models. We add supervision from high-level features of the large networks in training to help the small network better learn object representation. More specifically, we conduct a mimic method for the features sampled from the entire feature map and use a transform layer to map features from the small network onto the same dimension of the large network. In training the small network, we optimize the similarity between features sampled from the same region on the feature maps of both networks. Extensive experiments are conducted on pedestrian and common object detection tasks using VGG, Inception and ResNet. On both Caltech and Pascal VOC, we show that the modified 2.5× accelerated Inception network achieves competitive performance as the full Inception Network. Our faster model runs at 80 FPS for a 1000×1500 large input with only a minor degradation of performance on Caltech.",
"In light of the powerful learning capability of deep neural networks (DNNs), deep (convolutional) models have been built in recent years to address the task of salient object detection. Although training such deep saliency models can significantly improve the detection performance, it requires large-scale manual supervision in the form of pixel-level human annotation, which is highly labor-intensive and time-consuming. To address this problem, this paper makes the earliest effort to train a deep salient object detector without using any human annotation. The key insight is “supervision by fusion”, i.e., generating useful supervisory signals from the fusion process of weak but fast unsupervised saliency models. Based on this insight, we combine an intra-image fusion stream and a inter-image fusion stream in the proposed framework to generate the learning curriculum and pseudo ground-truth for supervising the training of the deep salient object detector. Comprehensive experiments on four benchmark datasets demonstrate that our method can approach the same network trained with full supervision (within 2-5 performance gap) and, more encouragingly, even outperform a number of fully supervised state-of-the-art approaches."
]
} |
1901.10951 | 2914704920 | For autonomous vehicles to be able to operate successfully they need to be aware of other vehicles with sufficient time to make safe, stable plans. Given the possible closing speeds between two vehicles, this necessitates the ability to accurately detect distant vehicles. Many current image-based object detectors using convolutional neural networks exhibit excellent performance on existing datasets such as KITTI. However, the performance of these networks falls when detecting small (distant) objects. We demonstrate that incorporating radar data can boost performance in these difficult situations. We also introduce an efficient automated method for training data generation using cameras of different focal lengths. | A number of papers use automated methods for generating training labels. In @cite_1 , visual odometry from previous traversals is used to label driveable surfaces for semantic segmentation. Hoermann @cite_11 employs temporal consistency to generate labels by processing data both forwards and backwards in time. Recent work by Adhikari @cite_20 takes a labelling approach that is related to ours by also leveraging the power of an existing object detector to generate labelled data for a new task, although a small amount of manual labelling is still required. | {
"cite_N": [
"@cite_1",
"@cite_20",
"@cite_11"
],
"mid": [
"2029731618",
"2737129951",
"2152525669"
],
"abstract": [
"We address the task of learning a semantic segmentation from weakly supervised data. Our aim is to devise a system that predicts an object label for each pixel by making use of only image level labels during training – the information whether a certain object is present or not in the image. Such coarse tagging of images is faster and easier to obtain as opposed to the tedious task of pixelwise labeling required in state of the art systems. We cast this task naturally as a multiple instance learning (MIL) problem. We use Semantic Texton Forest (STF) as the basic framework and extend it for the MIL setting. We make use of multitask learning (MTL) to regularize our solution. Here, an external task of geometric context estimation is used to improve on the task of semantic segmentation. We report experimental results on the MSRC21 and the very challenging VOC2007 datasets. On MSRC21 dataset we are able, by using 276 weakly labeled images, to achieve the performance of a supervised STF trained on pixelwise labeled training set of 56 images, which is a significant reduction in supervision needed.",
"This paper deals with semantic segmentation of high-resolution (aerial) images where a semantic class label is assigned to each pixel via supervised classification as a basis for automatic map generation. Recently, deep convolutional neural networks (CNNs) have shown impressive performance and have quickly become the de-facto standard for semantic segmentation, with the added benefit that task-specific feature design is no longer necessary. However, a major downside of deep learning methods is that they are extremely data hungry, thus aggravating the perennial bottleneck of supervised classification, to obtain enough annotated training data. On the other hand, it has been observed that they are rather robust against noise in the training labels. This opens up the intriguing possibility to avoid annotating huge amounts of training data, and instead train the classifier from existing legacy data or crowd-sourced maps that can exhibit high levels of noise. The question addressed in this paper is: can training with large-scale publicly available labels replace a substantial part of the manual labeling effort and still achieve sufficient performance? Such data will inevitably contain a significant portion of errors, but in return virtually unlimited quantities of it are available in larger parts of the world. We adapt a state-of-the-art CNN architecture for semantic segmentation of buildings and roads in aerial images, and compare its performance when using different training data sets, ranging from manually labeled pixel-accurate ground truth of the same city to automatic training data derived from OpenStreetMap data from distant locations. We report our results that indicate that satisfying performance can be obtained with significantly less manual annotation effort, by exploiting noisy large-scale training data.",
"This paper presents a novel and robust approach to consistent labeling for people surveillance in multicamera systems. A general framework scalable to any number of cameras with overlapped views is devised. An offline training process automatically computes ground-plane homography and recovers epipolar geometry. When a new object is detected in any one camera, hypotheses for potential matching objects in the other cameras are established. Each of the hypotheses is evaluated using a prior and likelihood value. The prior accounts for the positions of the potential matching objects, while the likelihood is computed by warping the vertical axis of the new object on the field of view of the other cameras and measuring the amount of match. In the likelihood, two contributions (forward and backward) are considered so as to correctly handle the case of groups of people merged into single objects. Eventually, a maximum-a-posteriori approach estimates the best label assignment for the new object. Comparisons with other methods based on homography and extensive outdoor experiments demonstrate that the proposed approach is accurate and robust in coping with segmentation errors and in disambiguating groups."
]
} |
1901.10951 | 2914704920 | For autonomous vehicles to be able to operate successfully they need to be aware of other vehicles with sufficient time to make safe, stable plans. Given the possible closing speeds between two vehicles, this necessitates the ability to accurately detect distant vehicles. Many current image-based object detectors using convolutional neural networks exhibit excellent performance on existing datasets such as KITTI. However, the performance of these networks falls when detecting small (distant) objects. We demonstrate that incorporating radar data can boost performance in these difficult situations. We also introduce an efficient automated method for training data generation using cameras of different focal lengths. | Multi-modal object detection has been investigated in a number of works. In @cite_4 camera images are combined with both frontal and birds eye LIDAR views for 3D object detection. Data from cameras, LIDAR and radar are all fused in @cite_19 . | {
"cite_N": [
"@cite_19",
"@cite_4"
],
"mid": [
"2950952351",
"2155338019"
],
"abstract": [
"This paper aims at high-accuracy 3D object detection in autonomous driving scenario. We propose Multi-View 3D networks (MV3D), a sensory-fusion framework that takes both LIDAR point cloud and RGB images as input and predicts oriented 3D bounding boxes. We encode the sparse 3D point cloud with a compact multi-view representation. The network is composed of two subnetworks: one for 3D object proposal generation and another for multi-view feature fusion. The proposal network generates 3D candidate boxes efficiently from the bird's eye view representation of 3D point cloud. We design a deep fusion scheme to combine region-wise features from multiple views and enable interactions between intermediate layers of different paths. Experiments on the challenging KITTI benchmark show that our approach outperforms the state-of-the-art by around 25 and 30 AP on the tasks of 3D localization and 3D detection. In addition, for 2D detection, our approach obtains 10.3 higher AP than the state-of-the-art on the hard data among the LIDAR-based methods.",
"In this paper, we present a unifying approach for learning and recognition of objects in unstructured environments through exploration. Taking inspiration from how young infants learn objects, we establish four principles for object learning. First, early object detection is based on an attention mechanism detecting salient parts in the scene. Second, motion of the object allows more accurate object localization. Next, acquiring multiple observations of the object through manipulation allows a more robust representation of the object. And last, object recognition benefits from a multi-modal representation. Using these principles, we developed a unifying method including visual attention, smooth pursuit of the object, and a multi-view and multi-modal object representation. Our results indicate the effectiveness of this approach and the improvement of the system when multiple observations are acquired from active object manipulation."
]
} |
1901.11117 | 2952355681 | Recent works have highlighted the strength of the Transformer architecture on sequence tasks while, at the same time, neural architecture search (NAS) has begun to outperform human-designed models. Our goal is to apply NAS to search for a better alternative to the Transformer. We first construct a large search space inspired by the recent advances in feed-forward sequence models and then run evolutionary architecture search with warm starting by seeding our initial population with the Transformer. To directly search on the computationally expensive WMT 2014 English-German translation task, we develop the Progressive Dynamic Hurdles method, which allows us to dynamically allocate more resources to more promising candidate models. The architecture found in our experiments -- the Evolved Transformer -- demonstrates consistent improvement over the Transformer on four well-established language tasks: WMT 2014 English-German, WMT 2014 English-French, WMT 2014 English-Czech and LM1B. At a big model size, the Evolved Transformer establishes a new state-of-the-art BLEU score of 29.8 on WMT'14 English-German; at smaller sizes, it achieves the same quality as the original "big" Transformer with 37.6 less parameters and outperforms the Transformer by 0.7 BLEU at a mobile-friendly model size of 7M parameters. | RNNs have long been used as the default option for applying neural networks to sequence modeling @cite_16 @cite_31 , with LSTM @cite_10 and GRU @cite_20 architectures being the most popular. However, recent work has shown that RNNs are not necessary to build state-of-the-art sequence models. For example, many high performance convolutional models have been designed, such as WaveNet @cite_3 , Gated Convolution Networks @cite_34 , Conv Seq2Seq @cite_46 and Dynamic Lightweight Convolution model @cite_23 . Perhaps the most promising architecture in this direction is the Transformer architecture @cite_39 , which relies only on multi-head attention to convey spatial information. In this work, we use both convolutions and attention in our search space to leverage the strengths of both layer types. | {
"cite_N": [
"@cite_46",
"@cite_3",
"@cite_34",
"@cite_39",
"@cite_23",
"@cite_31",
"@cite_16",
"@cite_10",
"@cite_20"
],
"mid": [
"1922658220",
"196214544",
"2963196092",
"2174479785",
"2791366550",
"2953188482",
"2866343820",
"2774989306",
"2962889000"
],
"abstract": [
"In existing convolutional neural networks (CNNs), both convolution and pooling are locally performed for image regions separately, no contextual dependencies between different image regions have been taken into consideration. Such dependencies represent useful spatial structure information in images. Whereas recurrent neural networks (RNNs) are designed for learning contextual dependencies among sequential data by using the recurrent (feedback) connections. In this work, we propose the convolutional recurrent neural network (C-RNN), which learns the spatial dependencies between image regions to enhance the discriminative power of image representation. The C-RNN is trained in an end-to-end manner from raw pixel images. CNN layers are firstly processed to generate middle level features. RNN layer is then learned to encode spatial dependencies. The C-RNN can learn better image representation, especially for images with obvious spatial contextual dependencies. Our method achieves competitive performance on ILSVRC 2012, SUN 397, and MIT indoor.",
"Recurrent Neural Networks (RNNs) are very powerful sequence models that do not enjoy widespread use because it is extremely difficult to train them properly. Fortunately, recent advances in Hessian-free optimization have been able to overcome the difficulties associated with training RNNs, making it possible to apply them successfully to challenging sequence problems. In this paper we demonstrate the power of RNNs trained with the new Hessian-Free optimizer (HF) by applying them to character-level language modeling tasks. The standard RNN architecture, while effective, is not ideally suited for such tasks, so we introduce a new RNN variant that uses multiplicative (or \"gated\") connections which allow the current input character to determine the transition matrix from one hidden state vector to the next. After training the multiplicative RNN with the HF optimizer for five days on 8 high-end Graphics Processing Units, we were able to surpass the performance of the best previous single method for character-level language modeling – a hierarchical non-parametric sequence model. To our knowledge this represents the largest recurrent neural network application to date.",
"Recurrent Neural Networks (RNNs) with sophisticated units that implement a gating mechanism have emerged as powerful technique for modeling sequential signals such as speech or electroencephalography (EEG). The latter is the focus on this paper. A significant big data resource, known as the TUH EEG Corpus (TUEEG), has recently become available for EEG research, creating a unique opportunity to evaluate these recurrent units on the task of seizure detection. In this study, we compare two types of recurrent units: long short-term memory units (LSTM) and gated recurrent units (GRU). These are evaluated using a state of the art hybrid architecture that integrates Convolutional Neural Networks (CNNs) with RNNs. We also investigate a variety of initialization methods and show that initialization is crucial since poorly initialized networks cannot be trained. Furthermore, we explore regularization of these convolutional gated recurrent networks to address the problem of overfitting. Our experiments revealed that convolutional LSTM networks can achieve significantly better performance than convolutional GRU networks. The convolutional LSTM architecture with proper initialization and regularization delivers 30 sensitivity at 6 false alarms per 24 hours.",
"We integrate the recently proposed spatial transformer network (SPN) [Jaderberg et. al 2015] into a recurrent neural network (RNN) to form an RNN-SPN model. We use the RNN-SPN to classify digits in cluttered MNIST sequences. The proposed model achieves a single digit error of 1.5 compared to 2.9 for a convolutional networks and 2.0 for convolutional networks with SPN layers. The SPN outputs a zoomed, rotated and skewed version of the input image. We investigate different down-sampling factors (ratio of pixel in input and output) for the SPN and show that the RNN-SPN model is able to down-sample the input images without deteriorating performance. The down-sampling in RNN-SPN can be thought of as adaptive down-sampling that minimizes the information loss in the regions of interest. We attribute the superior performance of the RNN-SPN to the fact that it can attend to a sequence of regions of interest.",
"Recurrent neural networks (RNNs) have been widely used for processing sequential data. However, RNNs are commonly difficult to train due to the well-known gradient vanishing and exploding problems and hard to learn long-term patterns. Long short-term memory (LSTM) and gated recurrent unit (GRU) were developed to address these problems, but the use of hyperbolic tangent and the sigmoid action functions results in gradient decay over layers. Consequently, construction of an efficiently trainable deep network is challenging. In addition, all the neurons in an RNN layer are entangled together and their behaviour is hard to interpret. To address these problems, a new type of RNN, referred to as independently recurrent neural network (IndRNN), is proposed in this paper, where neurons in the same layer are independent of each other and they are connected across layers. We have shown that an IndRNN can be easily regulated to prevent the gradient exploding and vanishing problems while allowing the network to learn long-term dependencies. Moreover, an IndRNN can work with non-saturated activation functions such as relu (rectified linear unit) and be still trained robustly. Multiple IndRNNs can be stacked to construct a network that is deeper than the existing RNNs. Experimental results have shown that the proposed IndRNN is able to process very long sequences (over 5000 time steps), can be used to construct very deep networks (21 layers used in the experiment) and still be trained robustly. Better performances have been achieved on various tasks by using IndRNNs compared with the traditional RNN and LSTM.",
"Recurrent neural networks (RNNs) have shown clear superiority in sequence modeling, particularly the ones with gated units, such as long short-term memory (LSTM) and gated recurrent unit (GRU). However, the dynamic properties behind the remarkable performance remain unclear in many applications, e.g., automatic speech recognition (ASR). This paper employs visualization techniques to study the behavior of LSTM and GRU when performing speech recognition tasks. Our experiments show some interesting patterns in the gated memory, and some of them have inspired simple yet effective modifications on the network structure. We report two of such modifications: (1) lazy cell update in LSTM, and (2) shortcut connections for residual learning. Both modifications lead to more comprehensible and powerful networks.",
"Recurrent neural networks (RNNs) sequentially process data by updating their state with each new data point, and have long been the de facto choice for sequence modeling tasks. However, their inherently sequential computation makes them slow to train. Feed-forward and convolutional architectures have recently been shown to achieve superior results on some sequence modeling tasks such as machine translation, with the added advantage that they concurrently process all inputs in the sequence, leading to easy parallelization and faster training times. Despite these successes, however, popular feed-forward sequence models like the Transformer fail to generalize in many simple tasks that recurrent models handle with ease, e.g. copying strings or even simple logical inference when the string or formula lengths exceed those observed at training time. We propose the Universal Transformer (UT), a parallel-in-time self-attentive recurrent sequence model which can be cast as a generalization of the Transformer model and which addresses these issues. UTs combine the parallelizability and global receptive field of feed-forward sequence models like the Transformer with the recurrent inductive bias of RNNs. We also add a dynamic per-position halting mechanism and find that it improves accuracy on several tasks. In contrast to the standard Transformer, under certain assumptions, UTs can be shown to be Turing-complete. Our experiments show that UTs outperform standard Transformers on a wide range of algorithmic and language understanding tasks, including the challenging LAMBADA language modeling task where UTs achieve a new state of the art, and machine translation where UTs achieve a 0.9 BLEU improvement over Transformers on the WMT14 En-De dataset.",
"Convolutional neural networks (CNNs) have demonstrated their ability object detection of very high resolution remote sensing images. However, CNNs have obvious limitations for modeling geometric variations in remote sensing targets. In this paper, we introduced a CNN structure, namely deformable ConvNet, to address geometric modeling in object recognition. By adding offsets to the convolution layers, feature mapping of CNN can be applied to unfixed locations, enhancing CNNs’ visual appearance understanding. In our work, a deformable region-based fully convolutional networks (R-FCN) was constructed by substituting the regular convolution layer with a deformable convolution layer. To efficiently use this deformable convolutional neural network (ConvNet), a training mechanism is developed in our work. We first set the pre-trained R-FCN natural image model as the default network parameters in deformable R-FCN. Then, this deformable ConvNet was fine-tuned on very high resolution (VHR) remote sensing images. To remedy the increase in lines like false region proposals, we developed aspect ratio constrained non maximum suppression (arcNMS). The precision of deformable ConvNet for detecting objects was then improved. An end-to-end approach was then developed by combining deformable R-FCN, a smart fine-tuning strategy and aspect ratio constrained NMS. The developed method was better than a state-of-the-art benchmark in object detection without data augmentation.",
"Abstract Recurrent Neural Networks (RNNs) have been widely used in natural language processing and computer vision. Amongst them, the Hierarchical Multi-scale RNN (HM-RNN), a recently proposed multi-scale hierarchical RNN, can automatically learn the hierarchical temporal structure from data. In this paper, we extend the work to solve the computer vision task of action recognition. However, in sequence-to-sequence models like RNN, it is normally very hard to discover the relationships between inputs and outputs given static inputs. As a solution, the attention mechanism can be applied to extract the relevant information from the inputs thus facilitating the modeling of the input–output relationships. Based on these considerations, we propose a novel attention network, namely Hierarchical Multi-scale Attention Network (HM-AN), by incorporating the attention mechanism into the HM-RNN and applying it to action recognition. A newly proposed gradient estimation method for stochastic neurons, namely Gumbel-softmax, is exploited to implement the temporal boundary detectors and the stochastic hard attention mechanism. To reduce the negative effect of the temperature sensitivity of the Gumbel-softmax, an adaptive temperature training method is applied to improve the system performance. The experimental results demonstrate the improved effect of HM-AN over LSTM with attention on the vision task. Through visualization of what has been learnt by the network, it can be observed that both the attention regions of the images and the hierarchical temporal structure can be captured by a HM-AN."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | Go-Explore is reminiscent of earlier work that separates exploration and exploitation (e.g. ), in which exploration follows a reward-agnostic Goal Exploration Process @cite_42 (an algorithm similar to novelty search @cite_46 ), from which experience is collected to prefill the replay buffer of an off-policy RL algorithm, in this case DDPG @cite_66 . This algorithm then extracts the highest-rewarding policy from the experience gathered. In contrast, Go-Explore further decomposes exploration into three elements: Accumulate stepping stones (interestingly different states), return to promising stepping stones, and explore from them in search of additional stepping stones (i.e. principles 1 and 2 above). The impressive results Go-Explore achieves by slotting in very simple algorithms for each element shows the value of this decomposition. | {
"cite_N": [
"@cite_46",
"@cite_42",
"@cite_66"
],
"mid": [
"2914261249",
"2952720101",
"2963176272"
],
"abstract": [
"A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of \"superhuman\" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics).",
"In environments with uncertain dynamics exploration is necessary to learn how to perform well. Existing reinforcement learning algorithms provide strong exploration guarantees, but they tend to rely on an ergodicity assumption. The essence of ergodicity is that any state is eventually reachable from any other state by following a suitable policy. This assumption allows for exploration algorithms that operate by simply favoring states that have rarely been visited before. For most physical systems this assumption is impractical as the systems would break before any reasonable exploration has taken place, i.e., most physical systems don't satisfy the ergodicity assumption. In this paper we address the need for safe exploration methods in Markov decision processes. We first propose a general formulation of safety through ergodicity. We show that imposing safety by restricting attention to the resulting set of guaranteed safe policies is NP-hard. We then present an efficient algorithm for guaranteed safe, but potentially suboptimal, exploration. At the core is an optimization formulation in which the constraints restrict attention to a subset of the guaranteed safe policies and the objective favors exploration policies. Our framework is compatible with the majority of previously proposed exploration methods, which rely on an exploration bonus. Our experiments, which include a Martian terrain exploration problem, show that our method is able to explore better than classical exploration methods.",
"Exploration is a fundamental challenge in reinforcement learning (RL). Many current exploration methods for deep RL use task-agnostic objectives, such as information gain or bonuses based on state visitation. However, many practical applications of RL involve learning more than a single task, and prior tasks can be used to inform how exploration should be performed in new tasks. In this work, we study how prior tasks can inform an agent about how to explore effectively in new situations. We introduce a novel gradient-based fast adaptation algorithm -- model agnostic exploration with structured noise (MAESN) -- to learn exploration strategies from prior experience. The prior experience is used both to initialize a policy and to acquire a latent exploration space that can inject structured stochasticity into a policy, producing exploration strategies that are informed by prior knowledge and are more effective than random action-space noise. We show that MAESN is more effective at learning exploration strategies when compared to prior meta-RL methods, RL without learned exploration strategies, and task-agnostic exploration methods. We evaluate our method on a variety of simulated tasks: locomotion with a wheeled robot, locomotion with a quadrupedal walker, and object manipulation."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | The aspect of Go-Explore of first finding a solution and then robustifying around it has precedent in Guided Policy Search @cite_19 . However, this method requires a non-deceptive, non-sparse, differentiable loss function to find solutions, meaning it cannot be applied directly to problems where rewards are discrete, sparse, or deceptive, as both Atari and many real-world problems are. Further, Guided Policy Search requires having a differentiable model of the world or learning a set of local models, which to be tractable requires the full state of the system to be observable during training time. | {
"cite_N": [
"@cite_19"
],
"mid": [
"2104733512"
],
"abstract": [
"Direct policy search can effectively scale to high-dimensional systems, but complex policies with hundreds of parameters often present a challenge for such methods, requiring numerous samples and often falling into poor local optima. We present a guided policy search algorithm that uses trajectory optimization to direct policy learning and avoid poor local optima. We show how differential dynamic programming can be used to generate suitable guiding samples, and describe a regularized importance sampled policy optimization that incorporates these samples into the policy search. We evaluate the method by learning neural network controllers for planar swimming, hopping, and walking, as well as simulated 3D humanoid running."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | The idea of planning (searching in a deterministic model of the world to find a good strategy) and then training a policy to mimic what was learned is reminiscent of . It plans (in the Atari emulator) with UCT @cite_100 @cite_22 @cite_73 , which is slow, and then trains a much faster policy with supervised learning to imitate the planning algorithm. At first glance it seems that in UCT serves a similar role to the exploration phase in Go-Explore, but UCT is quite different in several ways that make it inferior for domains that are either high-dimensional or hard-exploration. That is true even though UCT does have a form of exploration bonus. | {
"cite_N": [
"@cite_100",
"@cite_73",
"@cite_22"
],
"mid": [
"2401523698",
"2914261249",
"2169498096"
],
"abstract": [
"The Atari 2600 games supported in the Arcade Learning Environment [, 2013] all feature a known initial (RAM) state and actions that have deterministic effects. Classical planners, however, cannot be used off-the-shelf as there is no compact PDDL-model of the games, and action effects and goals are not known a priori. Indeed, there are no explicit goals, and the planner must select actions on-line while interacting with a simulator that returns successor states and rewards. None of this precludes the use of blind lookahead algorithms for action selection like breadth-first search or Dijkstra's yet such methods are not effective over large state spaces. We thus turn to a different class of planning methods introduced recently that have been shown to be effective for solving large planning problems but which do not require prior knowledge of state transitions, costs (rewards) or goals. The empirical results over 54 Atari games show that the simplest such algorithm performs at the level of UCT, the state-of-the-art planning method in this domain, and suggest the potential of width-based methods for planning with simulators when factored, compact action models are not available.",
"A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of \"superhuman\" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics).",
"Imitation learning of sequential, goal-directed behavior by standard supervised techniques is often difficult. We frame learning such behaviors as a maximum margin structured prediction problem over a space of policies. In this approach, we learn mappings from features to cost so an optimal policy in an MDP with these cost mimics the expert's behavior. Further, we demonstrate a simple, provably efficient approach to structured maximum margin learning, based on the subgradient method, that leverages existing fast algorithms for inference. Although the technique is general, it is particularly relevant in problems where A* and dynamic programming approaches make learning policies tractable in problems beyond the limitations of a QP formulation. We demonstrate our approach applied to route planning for outdoor mobile robots, where the behavior a designer wishes a planner to execute is often clear, while specifying cost functions that engender this behavior is a much more difficult task."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | UCT plans in a model of the world so as to decide on the next action to take in the real environment. An exploration bonus is used during the planning phase, but only extrinsic rewards are considered when choosing the next action to take. This approach can improve performance in domains with relatively dense rewards, but fails in sparse rewards domains as rewards are likely to be beyond the planning horizon of the algorithm. Once planning what to do from one state is done, an action is taken and the planning process is run again from the next state. UCT does not try to explore all states, and each run of UCT is independent of which states were visited in previous planning steps. As such, UCT (either within an episode, or across episodes) does not try to discover new terrain: instead its exploration bonus only helps it within the current short-horizon planning phase. As mentioned in , UCT scores 0 on Montezuma's Revenge and Pitfall @cite_91 @cite_60 . | {
"cite_N": [
"@cite_91",
"@cite_60"
],
"mid": [
"91160652",
"1625390266"
],
"abstract": [
"Monte-Carlo planning algorithms, such as UCT, select actions at each decision epoch by intelligently expanding a single search tree given the available time and then selecting the best root action. Recent work has provided evidence that it can be advantageous to instead construct an ensemble of search trees and to make a decision according to a weighted vote. However, these prior investigations have only considered the application domains of Go and Solitaire and were limited in the scope of ensemble configurations considered. In this paper, we conduct a more exhaustive empirical study of ensemble Monte-Carlo planning using the UCT algorithm in a set of six additional domains. In particular, we evaluate the advantages of a broad set of ensemble configurations in terms of space and time efficiency in both parallel and single-core models. Our results demonstrate that ensembles are an effective way to improve performance per unit time given a parallel time model and performance per unit space in a single-core model. However, contrary to prior isolated observations, we did not find significant evidence that ensembles improve performance per unit time in a single-core model.",
"For large state-space Markovian Decision Problems Monte-Carlo planning is one of the few viable approaches to find near-optimal solutions. In this paper we introduce a new algorithm, UCT, that applies bandit ideas to guide Monte-Carlo planning. In finite-horizon or discounted MDPs the algorithm is shown to be consistent and finite sample bounds are derived on the estimation error due to sampling. Experimental results show that in several domains, UCT is significantly more efficient than its alternatives."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | Another approach to planning is Fractal Monte Carlo (FMC) @cite_78 . When choosing the next action, it takes into account both the expected reward and novelty of that action, and in that way is more similar to Go-Explore. In FMC, a planning process is initiated from each state the agent visits. Planning is done within a deterministic version of the game emulator. A fixed number of workers are started in the state from which planning is occurring, and they perform random walks in state space. Periodically, workers that have accumulated lower reward and or are in less novel states are replaced by clones'' of more successful workers. Novelty is approximated as the Euclidean distance of the worker's state (in the original, raw, observation space) to that of a randomly selected other worker. | {
"cite_N": [
"@cite_78"
],
"mid": [
"1826354755"
],
"abstract": [
"We investigate the problem of planning under uncertainty, with application to mobile robotics. We propose a probabilistic framework in which the robot bases its decisions on the generalized belief, which is a probabilistic description of its own state and of external variables of interest. The approach naturally leads to a dual-layer architecture: an inner estimation layer, which performs inference to predict the outcome of possible decisions; and an outer decisional layer which is in charge of deciding the best action to undertake. Decision making is entrusted to a model predictive control MPC scheme. The formulation is valid for general cost functions and does not discretize the state or control space, enabling planning in continuous domain. Moreover, it allows to relax the assumption of maximum likelihood observations: predicted measurements are treated as random variables, and binary random variables are used to model the event that a measurement is actually taken by the robot. We successfully apply our approach to the problem of uncertainty-constrained exploration, in which the robot has to perform tasks in an unknown environment, while maintaining localization uncertainty within given bounds. We present an extensive numerical analysis of the proposed approach and compare it against related work. In practice, our planning approach produces smooth and natural trajectories and is able to impose soft upper bounds on the uncertainty. Finally, we exploit the results of this analysis to identify current limitations and show that the proposed framework can accommodate several desirable extensions."
]
} |
1901.10995 | 2914261249 | A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics). | On Pitfall, SOORL @cite_51 was the first planning algorithm to achieve a non-zero score, but did so in a deterministic test environment. It does so through a combination of learning a model of the environment, domain knowledge, and a value function that is optimistic about the value of unseen states, thus effectively providing an exploration bonus. At the end of 50 episodes of training, which was the maximum reported number of episodes, SOORL achieves an average of about 200 points across runs, and its best run scored an average of 606.6 with a maximum of 4,000. | {
"cite_N": [
"@cite_51"
],
"mid": [
"2111764152"
],
"abstract": [
"Most provably-efficient reinforcement learning algorithms introduce optimism about poorly-understood states and actions to encourage exploration. We study an alternative approach for efficient exploration: posterior sampling for reinforcement learning (PSRL). This algorithm proceeds in repeated episodes of known duration. At the start of each episode, PSRL updates a prior distribution over Markov decision processes and takes one sample from this posterior. PSRL then follows the policy that is optimal for this sample during the episode. The algorithm is conceptually simple, computationally efficient and allows an agent to encode prior knowledge in a natural way. We establish an O(τS √AT) bound on expected regret, where T is time, τ is the episode length and S and A are the cardinalities of the state and action spaces. This bound is one of the first for an algorithm not based on optimism, and close to the state of the art for any reinforcement learning algorithm. We show through simulation that PSRL significantly outperforms existing algorithms with similar regret bounds."
]
} |
1901.10912 | 2914607694 | We propose to meta-learn causal structures based on how fast a learner adapts to new distributions arising from sparse distributional changes, e.g. due to interventions, actions of agents and other sources of non-stationarities. We show that under this assumption, the correct causal structural choices lead to faster adaptation to modified distributions because the changes are concentrated in one or just a few mechanisms when the learned knowledge is modularized appropriately. This leads to sparse expected gradients and a lower effective number of degrees of freedom needing to be relearned while adapting to the change. It motivates using the speed of adaptation to a modified distribution as a meta-learning objective. We demonstrate how this can be used to determine the cause-effect relationship between two observed variables. The distributional changes do not need to correspond to standard interventions (clamping a variable), and the learner has no direct knowledge of these interventions. We show that causal structures can be parameterized via continuous variables and learned end-to-end. We then explore how these ideas could be used to also learn an encoder that would map low-level observed variables to unobserved causal variables leading to faster adaptation out-of-distribution, learning a representation space where one can satisfy the assumptions of independent mechanisms and of small and sparse changes in these mechanisms due to actions and non-stationarities. | Approaches for Bayesian network structure learning based on discrete search over model structures and simulated annealing are reviewed in . There, it has been common to use Minimum Description Length (MDL) principles to score and search over models , or the Bayesian Information Criterion (BIC) to search for models with high relative posterior probability @cite_2 . Prior work such as has also relied upon purely observational data, without the possibility of interventions and therefore focused on learning likelihood or hypothesis equivalence classes for network structures. Since then, numerous methods have also been devised to infer the causal direction from purely observational data , based on specific, generally parametric assumptions, on the underlying causal graph. Pearl's seminal work on do-calculus @cite_5 @cite_1 @cite_3 lays a foundation for expressing the impact of interventions on probabilistic graphical models -- we use it in our work. In contrast, here we are proposing a meta-learning objective function for learning causal structure, not requiring any specific constraints on causal graph structure, only on the sparsity of the changes in distribution in the correct causal graph parametrization. | {
"cite_N": [
"@cite_5",
"@cite_1",
"@cite_3",
"@cite_2"
],
"mid": [
"2170112109",
"2950873347",
"2144731007",
"1641347315"
],
"abstract": [
"We describe a Bayesian approach for learning Bayesian networks from a combination of prior knowledge and statistical data. First and foremost, we develop a methodology for assessing informative priors needed for learning. Our approach is derived from a set of assumptions made previously as well as the assumption of likelihood equivalence, which says that data should not help to discriminate network structures that represent the same assertions of conditional independence. We show that likelihood equivalence when combined with previously made assumptions implies that the user's priors for network parameters can be encoded in a single Bayesian network for the next case to be seen—a prior network—and a single measure of confidence for that network. Second, using these priors, we show how to compute the relative posterior probabilities of network structures given data. Third, we describe search methods for identifying network structures with high posterior probabilities. We describe polynomial algorithms for finding the highest-scoring network structures in the special case where every node has at most k e 1 parent. For the general case (k > 1), which is NP-hard, we review heuristic search algorithms including local search, iterative local search, and simulated annealing. Finally, we describe a methodology for evaluating Bayesian-network learning algorithms, and apply this approach to a comparison of various approaches.",
"We present a hybrid constraint-based Bayesian algorithm for learning causal networks in the presence of sparse data. The algorithm searches the space of equivalence classes of models (essential graphs) using a heuristic based on conventional constraint-based techniques. Each essential graph is then converted into a directed acyclic graph and scored using a Bayesian scoring metric. Two variants of the algorithm are developed and tested using data from randomly generated networks of sizes from 15 to 45 nodes with data sizes ranging from 250 to 2000 records. Both variations are compared to, and found to consistently outperform two variations of greedy search with restarts.",
"In many multivariate domains, we are interested in analyzing the dependency structure of the underlying distribution, e.g., whether two variables are in direct interaction. We can represent dependency structures using Bayesian network models. To analyze a given data set, Bayesian model selection attempts to find the most likely (MAP) model, and uses its structure to answer these questions. However, when the amount of available data is modest, there might be many models that have non-negligible posterior. Thus, we want compute the Bayesian posterior of a feature, i.e., the total posterior probability of all models that contain it. In this paper, we propose a new approach for this task. We first show how to efficiently compute a sum over the exponential number of networks that are consistent with a fixed order over network variables. This allows us to compute, for a given order, both the marginal probability of the data and the posterior of a feature. We then use this result as the basis for an algorithm that approximates the Bayesian posterior of a feature. Our approach uses a Markov Chain Monte Carlo (MCMC) method, but over orders rather than over network structures. The space of orders is smaller and more regular than the space of structures, and has much a smoother posterior “landscape”. We present empirical results on synthetic and real-life datasets that compare our approach to full model averaging (when possible), to MCMC over network structures, and to a non-Bayesian bootstrap approach.",
"Whereas acausal Bayesian networks represent probabilistic independence, causal Bayesian networks represent causal relationships. In this paper, we examine Bayesian methods for learning both types of networks. Bayesian methods for learning acausal networks are fairly well developed. These methods often employ assumptions to facilitate the construction of priors, including the assumptions of parameter independence, parameter modularity, and likelihood equivalence. We show that although these assumptions also can be appropriate for learning causal networks, we need additional assumptions in order to learn causal networks. We introduce two sufficient assumptions, called mechanism independence and component independence. We show that these new assumptions, when combined with parameter independence, parameter modularity, and likelihood equivalence, allow us to apply methods for learning acausal networks to learn causal networks."
]
} |
1907.10471 | 2962991329 | We present a new two-stage 3D object detection framework, named sparse-to-dense 3D Object Detector (STD). The first stage is a bottom-up proposal generation network that uses raw point cloud as input to generate accurate proposals by seeding each point with a new spherical anchor. It achieves a high recall with less computation compared with prior works. Then, PointsPool is applied for generating proposal features by transforming their interior point features from sparse expression to compact representation, which saves even more computation time. In box prediction, which is the second stage, we implement a parallel intersection-over-union (IoU) branch to increase awareness of localization accuracy, resulting in further improved performance. We conduct experiments on KITTI dataset, and evaluate our method in terms of 3D object and Bird's Eye View (BEV) detection. Our method outperforms other state-of-the-arts by a large margin, especially on the hard set, with inference speed more than 10 FPS. | There are several approaches to tackle semantic segmentation on point cloud. In @cite_5 , a projection function converts LiDAR points to a UV map, which is then classified by 2D semantic segmentation @cite_5 @cite_7 @cite_3 in pixel level. In @cite_11 @cite_21 , a multi-view-based function produces the segmentation mask. This method fuses information from different views. Other solutions, such as @cite_2 @cite_1 @cite_4 @cite_6 @cite_27 , segment point cloud from raw LiDAR data. They directly generate features on each point while keeping original structural information. A max-pooling method gathers the global feature. It is then concatenated with local feature for processing. | {
"cite_N": [
"@cite_4",
"@cite_7",
"@cite_21",
"@cite_1",
"@cite_3",
"@cite_6",
"@cite_27",
"@cite_2",
"@cite_5",
"@cite_11"
],
"mid": [
"800653105",
"2766577666",
"2963336905",
"2895472109",
"2771796597",
"2962912109",
"2033979122",
"2946747865",
"2809722021",
"1913356549"
],
"abstract": [
"In this paper we present a novel street scene semantic recognition framework, which takes advantage of 3D point clouds captured by a high-definition LiDAR laser scanner. An important problem in object recognition is the need for sufficient labeled training data to learn robust classifiers. In this paper we show how to significantly reduce the need for manually labeled training data by reduction of scene complexity using non-supervised ground and building segmentation. Our system first automatically segments grounds point cloud, this is because the ground connects almost all other objects and we will use a connect component based algorithm to oversegment the point clouds. Then, using binary range image processing building facades will be detected. Remained point cloud will grouped into voxels which are then transformed to super voxels. Local 3D features extracted from super voxels are classified by trained boosted decision trees and labeled with semantic classes e.g. tree, pedestrian, car, etc. The proposed method is evaluated both quantitatively and qualitatively on a challenging fixed-position Terrestrial Laser Scanning (TLS) Velodyne data set and two Mobile Laser Scanning (MLS), Paris-rue-Madam and NAVTEQ True databases. Robust scene parsing results are reported.",
"In this paper, we address semantic segmentation of road-objects from 3D LiDAR point clouds. In particular, we wish to detect and categorize instances of interest, such as cars, pedestrians and cyclists. We formulate this problem as a point- wise classification problem, and propose an end-to-end pipeline called SqueezeSeg based on convolutional neural networks (CNN): the CNN takes a transformed LiDAR point cloud as input and directly outputs a point-wise label map, which is then refined by a conditional random field (CRF) implemented as a recurrent layer. Instance-level labels are then obtained by conventional clustering algorithms. Our CNN model is trained on LiDAR point clouds from the KITTI dataset, and our point-wise segmentation labels are derived from 3D bounding boxes from KITTI. To obtain extra training data, we built a LiDAR simulator into Grand Theft Auto V (GTA-V), a popular video game, to synthesize large amounts of realistic training data. Our experiments show that SqueezeSeg achieves high accuracy with astonishingly fast and stable runtime (8.7 ms per frame), highly desirable for autonomous driving applications. Furthermore, additionally training on synthesized data boosts validation accuracy on real-world data. Our source code and synthesized data will be open-sourced.",
"Semantic parsing of large-scale 3D point clouds is an important research topic in computer vision and remote sensing fields. Most existing approaches utilize hand-crafted features for each modality independently and combine them in a heuristic manner. They often fail to consider the consistency and complementary information among features adequately, which makes them difficult to capture high-level semantic structures. The features learned by most of the current deep learning methods can obtain high-quality image classification results. However, these methods are hard to be applied to recognize 3D point clouds due to unorganized distribution and various point density of data. In this paper, we propose a 3DCNN-DQN-RNN method which fuses the 3D convolutional neural network (CNN), Deep Q-Network (DQN) and Residual recurrent neural network (RNN)for an efficient semantic parsing of large-scale 3D point clouds. In our method, an eye window under control of the 3D CNN and DQN can localize and segment the points of the object's class efficiently. The 3D CNN and Residual RNN further extract robust and discriminative features of the points in the eye window, and thus greatly enhance the parsing accuracy of large-scale point clouds. Our method provides an automatic process that maps the raw data to the classification results. It also integrates object localization, segmentation and classification into one framework. Experimental results demonstrate that the proposed method outperforms the state-of-the-art point cloud classification methods.",
"Semantic segmentation of 3D unstructured point clouds remains an open research problem. Recent works predict semantic labels of 3D points by virtue of neural networks but take limited context knowledge into consideration. In this paper, a novel end-to-end approach for unstructured point cloud semantic segmentation, named 3P-RNN, is proposed to exploit the inherent contextual features. First the efficient pointwise pyramid pooling module is investigated to capture local structures at various densities by taking multi-scale neighborhood into account. Then the two-direction hierarchical recurrent neural networks (RNNs) are utilized to explore long-range spatial dependencies. Each recurrent layer takes as input the local features derived from unrolled cells and sweeps the 3D space along two directions successively to integrate structure knowledge. On challenging indoor and outdoor 3D datasets, the proposed framework demonstrates robust performance superior to state-of-the-arts.",
"3D data such as point clouds and meshes are becoming more and more available. The goal of this paper is to obtain 3D object and scene classification and semantic segmentation. Because point clouds have irregular formats, most of the existing methods convert the 3D data into multiple 2D projection images or 3D voxel grids. These representations are suited as input of conventional CNNs but they either ignore the underlying 3D geometrical structure or are constrained by data sparsity and computational complexity. Therefore, recent methods encode the coordinates of each point cloud to certain high dimensional features to cover the 3D space. However, by their design, these models are not able to sufficiently capture the local patterns. In this paper, we propose a method that directly uses point clouds as input and exploits the implicit space partition of k-d tree structure to learn the local contextual information and aggregate features at different scales hierarchically. Extensive experiments on challenging benchmarks show that our proposed model properly captures the local patterns to provide discriminative point set features. For the task of 3D scene semantic segmentation, our method outperforms the state-of-the-art on the challenging Stanford Large-Scale 3D Indoor Spaces Dataset(S3DIS) by a large margin.",
"We address semantic segmentation of road-objects from 3D LiDAR point clouds. In particular, we wish to detect and categorize instances of interest, such as cars, pedestrians and cyclists. We formulate this problem as a point-wise classification problem, and propose an end-to-end pipeline called SqueezeSeg based on convolutional neural networks (CNN): the CNN takes a transformed LiDAR point cloud as input and directly outputs a point-wise label map, which is then refined by a conditional random field (CRF) implemented as a recurrent layer. Instance-level labels are then obtained by conventional clustering algorithms. Our CNN model is trained on LiDAR point clouds from the KITTI [1] dataset, and our point-wise segmentation labels are derived from 3D bounding boxes from KITTI. To obtain extra training data, we built a LiDAR simulator into Grand Theft Auto @math (GTA-V), a popular video game, to synthesize large amounts of realistic training data. Our experiments show that SqueezeSeg achieves high accuracy with astonishingly fast and stable runtime ( @math ms per frame), highly desirable for autonomous driving. Furthermore, additionally training on synthesized data boosts validation accuracy on real-world data. Our source code is open-source released111https: github.com BichenWuUCB SqueezeSeg. The paper is accompanied by a video222https: youtu.be Xyn5Zd31m6s containing a high level introduction and demonstrations of this work.",
"Dense semantic segmentation of 3D point clouds is a challenging task. Many approaches deal with 2D semantic segmentation and can obtain impressive results. With the availability of cheap RGB-D sensors the field of indoor semantic segmentation has seen a lot of progress. Still it remains unclear how to deal with 3D semantic segmentation in the best way. We propose a novel 2D-3D label transfer based on Bayesian updates and dense pairwise 3D Conditional Random Fields. This approach allows us to use 2D semantic segmentations to create a consistent 3D semantic reconstruction of indoor scenes. To this end, we also propose a fast 2D semantic segmentation approach based on Randomized Decision Forests. Furthermore, we show that it is not needed to obtain a semantic segmentation for every frame in a sequence in order to create accurate semantic 3D reconstructions. We evaluate our approach on both NYU Depth datasets and show that we can obtain a significant speed-up compared to other methods.",
"This paper proposes RIU-Net (for Range-Image U-Net), the adaptation of a popular semantic segmentation network for the semantic segmentation of a 3D LiDAR point cloud. The point cloud is turned into a 2D range-image by exploiting the topology of the sensor. This image is then used as input to a U-net. This architecture has already proved its efficiency for the task of semantic segmentation of medical images. We demonstrate how it can also be used for the accurate semantic segmentation of a 3D LiDAR point cloud and how it represents a valid bridge between image processing and 3D point cloud processing. Our model is trained on range-images built from KITTI 3D object detection dataset. Experiments show that RIU-Net, despite being very simple, offers results that are comparable to the state-of-the-art of range-image based methods. Finally, we demonstrate that this architecture is able to operate at 90fps on a single GPU, which enables deployment for real-time segmentation.",
"This paper proposes a novel methodology for LiDAR point cloud processing that takes advantage of the implicit topology of various LiDAR sensors to derive 2D images from the point cloud while bringing spatial structure to each point. The interest of such a methodology is then proved by addressing the problems of segmentation and disocclusion of mobile objects in 3D LiDAR scenes acquired via street-based Mobile Mapping Systems (MMS). Most of the existing lines of research tackle those problems directly in the 3D space. This work promotes an alternative approach by using this image representation of the 3D point cloud, taking advantage of the fact that the problem of disocclusion has been intensively studied in the 2D image processing community over the past decade. Using the image derived from the sensor data by exploiting the sensor topology, a semi-automatic segmentation procedure based on depth histograms is presented. Then, a variational image inpainting technique is introduced to reconstruct the areas that are occluded by objects. Experiments and validation on real data prove the effectiveness of this methodology both in terms of accuracy and speed.",
"We propose an algorithm for semantic segmentation based on 3D point clouds derived from ego-motion. We motivate five simple cues designed to model specific patterns of motion and 3D world structure that vary with object category. We introduce features that project the 3D cues back to the 2D image plane while modeling spatial layout and context. A randomized decision forest combines many such features to achieve a coherent 2D segmentation and recognize the object categories present. Our main contribution is to show how semantic segmentation is possible based solely on motion-derived 3D world structure. Our method works well on sparse, noisy point clouds, and unlike existing approaches, does not need appearance-based descriptors. Experiments were performed on a challenging new video database containing sequences filmed from a moving car in daylight and at dusk. The results confirm that indeed, accurate segmentation and recognition are possible using only motion and 3D world structure. Further, we show that the motion-derived information complements an existing state-of-the-art appearance-based method, improving both qualitative and quantitative performance."
]
} |
1907.10471 | 2962991329 | We present a new two-stage 3D object detection framework, named sparse-to-dense 3D Object Detector (STD). The first stage is a bottom-up proposal generation network that uses raw point cloud as input to generate accurate proposals by seeding each point with a new spherical anchor. It achieves a high recall with less computation compared with prior works. Then, PointsPool is applied for generating proposal features by transforming their interior point features from sparse expression to compact representation, which saves even more computation time. In box prediction, which is the second stage, we implement a parallel intersection-over-union (IoU) branch to increase awareness of localization accuracy, resulting in further improved performance. We conduct experiments on KITTI dataset, and evaluate our method in terms of 3D object and Bird's Eye View (BEV) detection. Our method outperforms other state-of-the-arts by a large margin, especially on the hard set, with inference speed more than 10 FPS. | For multi-view methods, MV3D @cite_0 projects LiDAR point cloud to BEV and trains a Region Proposal Network (RPN) to generate positive proposals. It merges features from BEV, image view and front view in order to generate refined 3D bounding boxes. AVOD @cite_18 improves MV3D by fusing image and BEV features like @cite_33 . Unlike MV3D, which only merges features in the refinement phase, it also merges features from multiple views in the RPN phase to generate positive proposals. These methods still have the limitation when detecting small objects such as pedestrians and cyclists. They do not deal with cases with multiple objects in depth direction. | {
"cite_N": [
"@cite_0",
"@cite_18",
"@cite_33"
],
"mid": [
"2950952351",
"2555618208",
"2774996270"
],
"abstract": [
"This paper aims at high-accuracy 3D object detection in autonomous driving scenario. We propose Multi-View 3D networks (MV3D), a sensory-fusion framework that takes both LIDAR point cloud and RGB images as input and predicts oriented 3D bounding boxes. We encode the sparse 3D point cloud with a compact multi-view representation. The network is composed of two subnetworks: one for 3D object proposal generation and another for multi-view feature fusion. The proposal network generates 3D candidate boxes efficiently from the bird's eye view representation of 3D point cloud. We design a deep fusion scheme to combine region-wise features from multiple views and enable interactions between intermediate layers of different paths. Experiments on the challenging KITTI benchmark show that our approach outperforms the state-of-the-art by around 25 and 30 AP on the tasks of 3D localization and 3D detection. In addition, for 2D detection, our approach obtains 10.3 higher AP than the state-of-the-art on the hard data among the LIDAR-based methods.",
"This paper aims at high-accuracy 3D object detection in autonomous driving scenario. We propose Multi-View 3D networks (MV3D), a sensory-fusion framework that takes both LIDAR point cloud and RGB images as input and predicts oriented 3D bounding boxes. We encode the sparse 3D point cloud with a compact multi-view representation. The network is composed of two subnetworks: one for 3D object proposal generation and another for multi-view feature fusion. The proposal network generates 3D candidate boxes efficiently from the birds eye view representation of 3D point cloud. We design a deep fusion scheme to combine region-wise features from multiple views and enable interactions between intermediate layers of different paths. Experiments on the challenging KITTI benchmark show that our approach outperforms the state-of-the-art by around 25 and 30 AP on the tasks of 3D localization and 3D detection. In addition, for 2D detection, our approach obtains 14.9 higher AP than the state-of-the-art on the hard data among the LIDAR-based methods.",
"We present AVOD, an Aggregate View Object Detection network for autonomous driving scenarios. The proposed neural network architecture uses LIDAR point clouds and RGB images to generate features that are shared by two subnetworks: a region proposal network (RPN) and a second stage detector network. The proposed RPN uses a novel architecture capable of performing multimodal feature fusion on high resolution feature maps to generate reliable 3D object proposals for multiple object classes in road scenes. Using these proposals, the second stage detection network performs accurate oriented 3D bounding box regression and category classification to predict the extents, orientation, and classification of objects in 3D space. Our proposed architecture is shown to produce state of the art results on the KITTI 3D object detection benchmark while running in real time with a low memory footprint, making it a suitable candidate for deployment on autonomous vehicles. Code is at: this https URL"
]
} |
1907.10599 | 2963790895 | Are neural networks biased toward simple functions? Does depth always help learn more complex features? Is training the last layer of a network as good as training all layers? These questions seem unrelated at face value, but in this work we give all of them a common treatment from the spectral perspective. We will study the spectra of the *Conjugate Kernel*, CK, (also called the *Neural Network-Gaussian Process Kernel*), and the *Neural Tangent Kernel*, NTK. Roughly, the CK and the NTK tell us respectively "what a network looks like at initialization"and "what a network looks like during and after training." Their spectra then encode valuable information about the initial distribution and the training and generalization properties of neural networks. By analyzing the eigenvalues, we lend novel insights into the questions put forth at the beginning, and we verify these insights by extensive experiments of neural networks. We believe the computational tools we develop here for analyzing the spectra of CK and NTK serve as a solid foundation for future studies of deep neural networks. We have open-sourced the code for it and for generating the plots in this paper at this http URL. | This signal propagation perspective can be refined via random matrix theory @cite_25 @cite_24 . In these works, free probability is leveraged to compute the singular value distribution of the input-output map given by the random neural network, as the input dimension and width tend to infinity together. Other works also investigate various questions of neural network training and generalization from the random matrix perspective . | {
"cite_N": [
"@cite_24",
"@cite_25"
],
"mid": [
"2752851182",
"2953338282"
],
"abstract": [
"Neural network configurations with random weights play an important role in the analysis of deep learning. They define the initial loss landscape and are closely related to kernel and random feature methods. Despite the fact that these networks are built out of random matrices, the vast and powerful machinery of random matrix theory has so far found limited success in studying them. A main obstacle in this direction is that neural networks are nonlinear, which prevents the straightforward utilization of many of the existing mathematical results. In this work, we open the door for direct applications of random matrix theory to deep learning by demonstrating that the pointwise nonlinearities typically applied in neural networks can be incorporated into a standard method of proof in random matrix theory known as the moments method. The test case for our study is the Gram matrix @math , @math , where @math is a random weight matrix, @math is a random data matrix, and @math is a pointwise nonlinear activation function. We derive an explicit representation for the trace of the resolvent of this matrix, which defines its limiting spectral distribution. We apply these results to the computation of the asymptotic performance of single-layer random feature methods on a memorization task and to the analysis of the eigenvalues of the data covariance matrix as it propagates through a neural network. As a byproduct of our analysis, we identify an intriguing new class of activation functions with favorable properties.",
"We study the compressed sensing reconstruction problem for a broad class of random, band-diagonal sensing matrices. This construction is inspired by the idea of spatial coupling in coding theory. As demonstrated heuristically and numerically by KrzakalaEtAl , message passing algorithms can effectively solve the reconstruction problem for spatially coupled measurements with undersampling rates close to the fraction of non-zero coordinates. We use an approximate message passing (AMP) algorithm and analyze it through the state evolution method. We give a rigorous proof that this approach is successful as soon as the undersampling rate @math exceeds the (upper) R 'enyi information dimension of the signal, @math . More precisely, for a sequence of signals of diverging dimension @math whose empirical distribution converges to @math , reconstruction is with high probability successful from @math measurements taken according to a band diagonal matrix. For sparse signals, i.e., sequences of dimension @math and @math non-zero entries, this implies reconstruction from @math measurements. For discrete' signals, i.e., signals whose coordinates take a fixed finite set of values, this implies reconstruction from @math measurements. The result is robust with respect to noise, does not apply uniquely to random signals, but requires the knowledge of the empirical distribution of the signal @math ."
]
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.