PyTorchKR Papers

최신 논문

총 4,618편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:1509.08745

Compression of Deep Neural Networks on the Fly

Thanks to their state-of-the-art performance, deep neural networks are increasingly used for object recognition. To achieve these results, they use millions of parameters to be trained. However, when targeting embedded applications the size of these models becomes problematic. As a consequence, their usage on smartphones or other resource limited devices is prohibited. In this paper we introduce a novel compression method for deep neural networks that is performed during the learning phase. It consists in adding an extra regularization term to the cost function of fully-connected layers. We combine this method with Product Quantization (PQ) of the trained weights for higher savings in storage consumption. We evaluate our method on two data sets (MNIST and CIFAR10), on which we achieve significantly larger compression rates than state-of-the-art methods.

arXiv:1603.09037

Vector Quantization for Machine Vision

This paper shows how to reduce the computational cost for a variety of common machine vision tasks by operating directly in the compressed domain, particularly in the context of hardware acceleration. Pyramid Vector Quantization (PVQ) is the compression technique of choice and its properties are exploited to simplify Support Vector Machines (SVM), Convolutional Neural Networks(CNNs), Histogram of Oriented Gradients (HOG) features, interest points matching and other algorithms.

arXiv:1604.01475

Learning A Deep $\ell_\infty$ Encoder for Hashing

We investigate the $\ell_\infty$-constrained representation which demonstrates robustness to quantization errors, utilizing the tool of deep learning. Based on the Alternating Direction Method of Multipliers (ADMM), we formulate the original convex minimization problem as a feed-forward neural network, named \textit{Deep $\ell_\infty$ Encoder}, by introducing the novel Bounded Linear Unit (BLU) neuron and modeling the Lagrange multipliers as network biases. Such a structural prior acts as an effective network regularization, and facilitates the model initialization. We then investigate the effective use of the proposed model in the application of hashing, by coupling the proposed encoders under a supervised pairwise loss, to develop a \textit{Deep Siamese $\ell_\infty$ Network}, which can be optimized from end to end. Extensive experiments demonstrate the impressive performances of the proposed model. We also provide an in-depth analysis of its behaviors against the competitors.

arXiv:1604.02631

Grid Based Nonlinear Filtering Revisited: Recursive Estimation & Asymptotic Optimality

We revisit the development of grid based recursive approximate filtering of general Markov processes in discrete time, partially observed in conditionally Gaussian noise. The grid based filters considered rely on two types of state quantization: The \textit{Markovian} type and the \textit{marginal} type. We propose a set of novel, relaxed sufficient conditions, ensuring strong and fully characterized pathwise convergence of these filters to the respective MMSE state estimator. In particular, for marginal state quantizations, we introduce the notion of \textit{conditional regularity of stochastic kernels}, which, to the best of our knowledge, constitutes the most relaxed condition proposed, under which asymptotic optimality of the respective grid based filters is guaranteed. Further, we extend our convergence results, including filtering of bounded and continuous functionals of the state, as well as recursive approximate state prediction. For both Markovian and marginal quantizations, the whole development of the respective grid based filters relies more on linear-algebraic techniques and less on measure theoretic arguments, making the presentation considerably shorter and technically simpler.

arXiv:1601.00238

Dimensionality-Dependent Generalization Bounds for $k$-Dimensional Coding Schemes

The $k$-dimensional coding schemes refer to a collection of methods that attempt to represent data using a set of representative $k$-dimensional vectors, and include non-negative matrix factorization, dictionary learning, sparse coding, $k$-means clustering and vector quantization as special cases. Previous generalization bounds for the reconstruction error of the $k$-dimensional coding schemes are mainly dimensionality independent. A major advantage of these bounds is that they can be used to analyze the generalization error when data is mapped into an infinite- or high-dimensional feature space. However, many applications use finite-dimensional data features. Can we obtain dimensionality-dependent generalization bounds for $k$-dimensional coding schemes that are tighter than dimensionality-independent bounds when data is in a finite-dimensional feature space? The answer is positive. In this paper, we address this problem and derive a dimensionality-dependent generalization bound for $k$-dimensional coding schemes by bounding the covering number of the loss function class induced by the reconstruction error. The bound is of order $\mathcal{O}\left(\left(mk\ln(mkn)/n\right)^{\lambda_n}\right)$, where $m$ is the dimension of features, $k$ is the number of the columns in the linear implementation of coding schemes, $n$ is the size of sample, $\lambda_n>0.5$ when $n$ is finite and $\lambda_n=0.5$ when $n$ is infinite. We show that our bound can be tighter than previous results, because it avoids inducing the worst-case upper bound on $k$ of the loss function and converges faster. The proposed generalization bound is also applied to some specific coding schemes to demonstrate that the dimensionality-dependent bound is an indispensable complement to these dimensionality-independent generalization bounds.

arXiv:1605.04859

Reducing the Model Order of Deep Neural Networks Using Information Theory

Deep neural networks are typically represented by a much larger number of parameters than shallow models, making them prohibitive for small footprint devices. Recent research shows that there is considerable redundancy in the parameter space of deep neural networks. In this paper, we propose a method to compress deep neural networks by using the Fisher Information metric, which we estimate through a stochastic optimization method that keeps track of second-order information in the network. We first remove unimportant parameters and then use non-uniform fixed point quantization to assign more bits to parameters with higher Fisher Information estimates. We evaluate our method on a classification task with a convolutional neural network trained on the MNIST data set. Experimental results show that our method outperforms existing methods for both network pruning and quantization.

arXiv:1605.04034

Transfer Hashing with Privileged Information

Most existing learning to hash methods assume that there are sufficient data, either labeled or unlabeled, on the domain of interest (i.e., the target domain) for training. However, this assumption cannot be satisfied in some real-world applications. To address this data sparsity issue in hashing, inspired by transfer learning, we propose a new framework named Transfer Hashing with Privileged Information (THPI). Specifically, we extend the standard learning to hash method, Iterative Quantization (ITQ), in a transfer learning manner, namely ITQ+. In ITQ+, a new slack function is learned from auxiliary data to approximate the quantization error in ITQ. We developed an alternating optimization approach to solve the resultant optimization problem for ITQ+. We further extend ITQ+ to LapITQ+ by utilizing the geometry structure among the auxiliary data for learning more precise binary codes in the target domain. Extensive experiments on several benchmark datasets verify the effectiveness of our proposed approaches through comparisons with several state-of-the-art baselines.

arXiv:1605.08856

A Channelized Binning Method for Extraction of Dominant Color Pixel Value

The Color is one of the most important and easily identifiable features for describing the visual content. The MPEG standard has developed a number of descriptors that covers different aspects of the visual content. The Dominant color descriptor is one of them. This paper proposes a channelized binning approach a novel method for extraction of the dominant color pixel value which is a variant of the dominant color descriptor. The Channelized binning method treats the problem as a statistical problem and tries to avoid color quantization and interpolation guessing of number and centroid of dominant colors. Channelized binning is an iterative approach which automatically estimates the number of dominant pixel values and their centroids. It operates on 24 bit full RGB color space, by considering one color channel at a time and hence avoiding the color quantization. Results show that the proposed method can successfully extract dominant color pixel values.

arXiv:1511.06393

Fixed Point Quantization of Deep Convolutional Networks

In recent years increasingly complex architectures for deep convolution networks (DCNs) have been proposed to boost the performance on image recognition tasks. However, the gains in performance have come at a cost of substantial increase in computation and model storage resources. Fixed point implementation of DCNs has the potential to alleviate some of these complexities and facilitate potential deployment on embedded hardware. In this paper, we propose a quantizer design for fixed point implementation of DCNs. We formulate and solve an optimization problem to identify optimal fixed point bit-width allocation across DCN layers. Our experiments show that in comparison to equal bit-width settings, the fixed point DCNs with optimized bit width allocation offer >20% reduction in the model size without any loss in accuracy on CIFAR-10 benchmark. We also demonstrate that fine-tuning can further enhance the accuracy of fixed point DCNs beyond that of the original floating point model. In doing so, we report a new state-of-the-art fixed point performance of 6.78% error-rate on CIFAR-10 benchmark.

arXiv:1606.01550

Pairwise Quantization

We consider the task of lossy compression of high-dimensional vectors through quantization. We propose the approach that learns quantization parameters by minimizing the distortion of scalar products and squared distances between pairs of points. This is in contrast to previous works that obtain these parameters through the minimization of the reconstruction error of individual points. The proposed approach proceeds by finding a linear transformation of the data that effectively reduces the minimization of the pairwise distortions to the minimization of individual reconstruction errors. After such transformation, any of the previously-proposed quantization approaches can be used. Despite the simplicity of this transformation, the experiments demonstrate that it achieves considerable reduction of the pairwise distortions compared to applying quantization directly to the untransformed data.

arXiv:1606.00930

Comparison of 14 different families of classification algorithms on 115 binary datasets

We tested 14 very different classification algorithms (random forest, gradient boosting machines, SVM - linear, polynomial, and RBF - 1-hidden-layer neural nets, extreme learning machines, k-nearest neighbors and a bagging of knn, naive Bayes, learning vector quantization, elastic net logistic regression, sparse linear discriminant analysis, and a boosting of linear classifiers) on 115 real life binary datasets. We followed the Demsar analysis and found that the three best classifiers (random forest, gbm and RBF SVM) are not significantly different from each other. We also discuss that a change of less then 0.0112 in the error rate should be considered as an irrelevant change, and used a Bayesian ANOVA analysis to conclude that with high probability the differences between these three classifiers is not of practical consequence. We also verified the execution time of "standard implementations" of these algorithms and concluded that RBF SVM is the fastest (significantly so) both in training time and in training plus testing time.

arXiv:1606.01981

Deep neural networks are robust to weight binarization and other non-linear distortions

Recent results show that deep neural networks achieve excellent performance even when, during training, weights are quantized and projected to a binary representation. Here, we show that this is just the tip of the iceberg: these same networks, during testing, also exhibit a remarkable robustness to distortions beyond quantization, including additive and multiplicative noise, and a class of non-linear projections where binarization is just a special case. To quantify this robustness, we show that one such network achieves 11% test error on CIFAR-10 even with 0.68 effective bits per weight. Furthermore, we find that a common training heuristic--namely, projecting quantized weights during backpropagation--can be altered (or even removed) and networks still achieve a base level of robustness during testing. Specifically, training with weight projections other than quantization also works, as does simply clipping the weights, both of which have never been reported before. We confirm our results for CIFAR-10 and ImageNet datasets. Finally, drawing from these ideas, we propose a stochastic projection rule that leads to a new state of the art network with 7.64% test error on CIFAR-10 using no data augmentation.

arXiv:1605.09046

TripleSpin - a generic compact paradigm for fast machine learning computations

We present a generic compact computational framework relying on structured random matrices that can be applied to speed up several machine learning algorithms with almost no loss of accuracy. The applications include new fast LSH-based algorithms, efficient kernel computations via random feature maps, convex optimization algorithms, quantization techniques and many more. Certain models of the presented paradigm are even more compressible since they apply only bit matrices. This makes them suitable for deploying on mobile devices. All our findings come with strong theoretical guarantees. In particular, as a byproduct of the presented techniques and by using relatively new Berry-Esseen-type CLT for random vectors, we give the first theoretical guarantees for one of the most efficient existing LSH algorithms based on the $\textbf{HD}_{3}\textbf{HD}_{2}\textbf{HD}_{1}$ structured matrix ("Practical and Optimal LSH for Angular Distance"). These guarantees as well as theoretical results for other aforementioned applications follow from the same general theoretical principle that we present in the paper. Our structured family contains as special cases all previously considered structured schemes, including the recently introduced $P$-model. Experimental evaluation confirms the accuracy and efficiency of TripleSpin matrices.

arXiv:1512.07587

A Latent-Variable Lattice Model

Markov random field (MRF) learning is intractable, and its approximation algorithms are computationally expensive. We target a small subset of MRF that is used frequently in computer vision. We characterize this subset with three concepts: Lattice, Homogeneity, and Inertia; and design a non-markov model as an alternative. Our goal is robust learning from small datasets. Our learning algorithm uses vector quantization and, at time complexity O(U log U) for a dataset of U pixels, is much faster than that of general-purpose MRF.

arXiv:1606.06061

Fast, Compact, and High Quality LSTM-RNN Based Statistical Parametric Speech Synthesizers for Mobile Devices

Acoustic models based on long short-term memory recurrent neural networks (LSTM-RNNs) were applied to statistical parametric speech synthesis (SPSS) and showed significant improvements in naturalness and latency over those based on hidden Markov models (HMMs). This paper describes further optimizations of LSTM-RNN-based SPSS for deployment on mobile devices; weight quantization, multi-frame inference, and robust inference using an {\epsilon}-contaminated Gaussian loss function. Experimental results in subjective listening tests show that these optimizations can make LSTM-RNN-based SPSS comparable to HMM-based SPSS in runtime speed while maintaining naturalness. Evaluations between LSTM-RNN- based SPSS and HMM-driven unit selection speech synthesis are also presented.

arXiv:1607.01895

Random Walk Graph Laplacian based Smoothness Prior for Soft Decoding of JPEG Images

Given the prevalence of JPEG compressed images, optimizing image reconstruction from the compressed format remains an important problem. Instead of simply reconstructing a pixel block from the centers of indexed DCT coefficient quantization bins (hard decoding), soft decoding reconstructs a block by selecting appropriate coefficient values within the indexed bins with the help of signal priors. The challenge thus lies in how to define suitable priors and apply them effectively. In this paper, we combine three image priors---Laplacian prior for DCT coefficients, sparsity prior and graph-signal smoothness prior for image patches---to construct an efficient JPEG soft decoding algorithm. Specifically, we first use the Laplacian prior to compute a minimum mean square error (MMSE) initial solution for each code block. Next, we show that while the sparsity prior can reduce block artifacts, limiting the size of the over-complete dictionary (to lower computation) would lead to poor recovery of high DCT frequencies. To alleviate this problem, we design a new graph-signal smoothness prior (desired signal has mainly low graph frequencies) based on the left eigenvectors of the random walk graph Laplacian matrix (LERaG). Compared to previous graph-signal smoothness priors, LERaG has desirable image filtering properties with low computation overhead. We demonstrate how LERaG can facilitate recovery of high DCT frequencies of a piecewise smooth (PWS) signal via an interpretation of low graph frequency components as relaxed solutions to normalized cut in spectral clustering. Finally, we construct a soft decoding algorithm using the three signal priors with appropriate prior weights. Experimental results show that our proposal outperforms state-of-the-art soft decoding algorithms in both objective and subjective evaluations noticeably.

arXiv:1604.00697

A New Learning Method for Inference Accuracy, Core Occupation, and Performance Co-optimization on TrueNorth Chip

IBM TrueNorth chip uses digital spikes to perform neuromorphic computing and achieves ultrahigh execution parallelism and power efficiency. However, in TrueNorth chip, low quantization resolution of the synaptic weights and spikes significantly limits the inference (e.g., classification) accuracy of the deployed neural network model. Existing workaround, i.e., averaging the results over multiple copies instantiated in spatial and temporal domains, rapidly exhausts the hardware resources and slows down the computation. In this work, we propose a novel learning method on TrueNorth platform that constrains the random variance of each computation copy and reduces the number of needed copies. Compared to the existing learning method, our method can achieve up to 68.8% reduction of the required neuro-synaptic cores or 6.5X speedup, with even slightly improved inference accuracy.

arXiv:1607.08635

A 58.6mW Real-Time Programmable Object Detector with Multi-Scale Multi-Object Support Using Deformable Parts Model on 1920x1080 Video at 30fps

This paper presents a programmable, energy-efficient and real-time object detection accelerator using deformable parts models (DPM), with 2x higher accuracy than traditional rigid body models. With 8 deformable parts detection, three methods are used to address the high computational complexity: classification pruning for 33x fewer parts classification, vector quantization for 15x memory size reduction, and feature basis projection for 2x reduction of the cost of each classification. The chip is implemented in 65nm CMOS technology, and can process HD (1920x1080) images at 30fps without any off-chip storage while consuming only 58.6mW (0.94nJ/pixel, 1168 GOPS/W). The chip has two classification engines to simultaneously detect two different classes of objects. With a tested high throughput of 60fps, the classification engines can be time multiplexed to detect even more than two object classes. It is energy scalable by changing the pruning factor or disabling the parts classification.

arXiv:1604.06480

LOH and behold: Web-scale visual search, recommendation and clustering using Locally Optimized Hashing

We propose a novel hashing-based matching scheme, called Locally Optimized Hashing (LOH), based on a state-of-the-art quantization algorithm that can be used for efficient, large-scale search, recommendation, clustering, and deduplication. We show that matching with LOH only requires set intersections and summations to compute and so is easily implemented in generic distributed computing systems. We further show application of LOH to: a) large-scale search tasks where performance is on par with other state-of-the-art hashing approaches; b) large-scale recommendation where queries consisting of thousands of images can be used to generate accurate recommendations from collections of hundreds of millions of images; and c) efficient clustering with a graph-based algorithm that can be scaled to massive collections in a distributed environment or can be used for deduplication for small collections, like search results, performing better than traditional hashing approaches while only requiring a few milliseconds to run. In this paper we experiment on datasets of up to 100 million images, but in practice our system can scale to larger collections and can be used for other types of data that have a vector representation in a Euclidean space.

arXiv:1607.02720

Intra-layer Nonuniform Quantization for Deep Convolutional Neural Network

Deep convolutional neural network (DCNN) has achieved remarkable performance on object detection and speech recognition in recent years. However, the excellent performance of a DCNN incurs high computational complexity and large memory requirement. In this paper, an equal distance nonuniform quantization (ENQ) scheme and a K-means clustering nonuniform quantization (KNQ) scheme are proposed to reduce the required memory storage when low complexity hardware or software implementations are considered. For the VGG-16 and the AlexNet, the proposed nonuniform quantization schemes reduce the number of required memory storage by approximately 50\% while achieving almost the same or even better classification accuracy compared to the state-of-the-art quantization method. Compared to the ENQ scheme, the proposed KNQ scheme provides a better tradeoff when higher accuracy is required.