Adversarial examples (AEs) are images that can mislead deep neural network (DNN) classifiers via introducing slight perturbations into original images. Recent work has shown that detecting AEs can be more effective against AEs than preventing them from being generated. However, the state-of-the-art AE detection still shows a high false positive rate, thereby rejecting a considerable amount of normal images. To address this issue, we propose HAWKEYE, which is a separate neural network that analyzes the output layer of the DNN, and detects AEs. HAWKEYE's AE detector utilizes a quantized version of an input image as a reference, and is trained to distinguish the variation characteristics of the DNN output on an input image from the DNN output on its reference image. We also show that cascading our AE detectors that are trained for different quantization step sizes can drastically reduce a false positive rate, while keeping a detection rate high.
We propose methods to train convolutional neural networks (CNNs) with both binarized weights and activations, leading to quantized models that are specifically friendly to mobile devices with limited power capacity and computation resources. Previous works on quantizing CNNs often seek to approximate the floating-point information using a set of discrete values, which we call value approximation, typically assuming the same architecture as the full-precision networks. Here we take a novel "structure approximation" view of quantization -- it is very likely that different architectures designed for low-bit networks may be better for achieving good performance. In particular, we propose a "network decomposition" strategy, termed Group-Net, in which we divide the network into groups. Thus, each full-precision group can be effectively reconstructed by aggregating a set of homogeneous binary branches. In addition, we learn effective connections among groups to improve the representation capability. Moreover, the proposed Group-Net shows strong generalization to other tasks. For instance, we extend Group-Net for accurate semantic segmentation by embedding rich context into the binary structure. Furthermore, for the first time, we apply binary neural networks to object detection. Experiments on both classification, semantic segmentation and object detection tasks demonstrate the superior performance of the proposed methods over various quantized networks in the literature. Our methods outperform the previous best binary neural networks in terms of accuracy and computation efficiency.
Binary Neural Networks (BNNs) are difficult to train, and suffer from drop of accuracy. It appears in practice that BNNs fail to train in the absence of Batch Normalization (BatchNorm) layer. We find the main role of BatchNorm is to avoid exploding gradients in the case of BNNs. This finding suggests that the common initialization methods developed for full-precision networks are irrelevant to BNNs. We build a theoretical study on the role of BatchNorm in binary training, backed up by numerical experiments.
Binary optimization, a representative subclass of discrete optimization, plays an important role in mathematical optimization and has various applications in computer vision and machine learning. Usually, binary optimization problems are NP-hard and difficult to solve due to the binary constraints, especially when the number of variables is very large. Existing methods often suffer from high computational costs or large accumulated quantization errors, or are only designed for specific tasks. In this paper, we propose a fast algorithm to find effective approximate solutions for general binary optimization problems. The proposed algorithm iteratively solves minimization problems related to the linear surrogates of loss functions, which leads to the updating of some binary variables most impacting the value of loss functions in each step. Our method supports a wide class of empirical objective functions with/without restrictions on the numbers of $1$s and $-1$s in the binary variables. Furthermore, the theoretical convergence of our algorithm is proven, and the explicit convergence rates are derived, for objective functions with Lipschitz continuous gradients, which are commonly adopted in practice. Extensive experiments on several binary optimization tasks and large-scale datasets demonstrate the superiority of the proposed algorithm over several state-of-the-art methods in terms of both effectiveness and efficiency.
To improve the execution speed and efficiency of neural networks in embedded systems, it is crucial to decrease the model size and computational complexity. In addition to conventional compression techniques, e.g., weight pruning and quantization, removing unimportant activations can reduce the amount of data communication and the computation cost. Unlike weight parameters, the pattern of activations is directly related to input data and thereby changes dynamically. To regulate the dynamic activation sparsity (DAS), in this work, we propose a generic low-cost approach based on winners-take-all (WTA) dropout technique. The network enhanced by the proposed WTA dropout, namely \textit{DASNet}, features structured activation sparsity with an improved sparsity level. Compared to the static feature map pruning methods, DASNets provide better computation cost reduction. The WTA technique can be easily applied in deep neural networks without incurring additional training variables. More importantly, DASNet can be seamlessly integrated with other compression techniques, such as weight pruning and quantization, without compromising on accuracy. Our experiments on various networks and datasets present significant run-time speedups with negligible accuracy loss.
Deep Learning (DL) has recently achieved tremendous success. A variety of DL frameworks and platforms play a key role to catalyze such progress. However, the differences in architecture designs and implementations of existing frameworks and platforms bring new challenges for DL software development and deployment. Till now, there is no study on how various mainstream frameworks and platforms influence both DL software development and deployment in practice. To fill this gap, we take the first step towards understanding how the most widely-used DL frameworks and platforms support the DL software development and deployment. We conduct a systematic study on these frameworks and platforms by using two types of DNN architectures and three popular datasets. (1) For development process, we investigate the prediction accuracy under the same runtime training configuration or same model weights/biases. We also study the adversarial robustness of trained models by leveraging the existing adversarial attack techniques. The experimental results show that the computing differences across frameworks could result in an obvious prediction accuracy decline, which should draw the attention of DL developers. (2) For deployment process, we investigate the prediction accuracy and performance (refers to time cost and memory consumption) when the trained models are migrated/quantized from PC to real mobile devices and web browsers. The DL platform study unveils that the migration and quantization still suffer from compatibility and reliability issues. Meanwhile, we find several DL software bugs by using the results as a benchmark. We further validate the results through bug confirmation from stakeholders and industrial positive feedback to highlight the implications of our study. Through our study, we summarize practical guidelines, identify challenges and pinpoint new research directions.
Neural Machine Translation (NMT) is resource intensive. We design a quantization procedure to compress NMT models better for devices with limited hardware capability. Because most neural network parameters are near zero, we employ logarithmic quantization in lieu of fixed-point quantization. However, we find bias terms are less amenable to log quantization but note they comprise a tiny fraction of the model, so we leave them uncompressed. We also propose to use an error-feedback mechanism during retraining, to preserve the compressed model as a stale gradient. We empirically show that NMT models based on Transformer or RNN architecture can be compressed up to 4-bit precision without any noticeable quality degradation. Models can be compressed up to binary precision, albeit with lower quality. The RNN architecture seems to be more robust to quantization, compared to the Transformer.
Transformer based architectures have become de-facto models used for a range of Natural Language Processing tasks. In particular, the BERT based models achieved significant accuracy gain for GLUE tasks, CoNLL-03 and SQuAD. However, BERT based models have a prohibitive memory footprint and latency. As a result, deploying BERT based models in resource constrained environments has become a challenging task. In this work, we perform an extensive analysis of fine-tuned BERT models using second order Hessian information, and we use our results to propose a novel method for quantizing BERT models to ultra low precision. In particular, we propose a new group-wise quantization scheme, and we use a Hessian based mix-precision method to compress the model further. We extensively test our proposed method on BERT downstream tasks of SST-2, MNLI, CoNLL-03, and SQuAD. We can achieve comparable performance to baseline with at most $2.3\%$ performance degradation, even with ultra-low precision quantization down to 2 bits, corresponding up to $13\times$ compression of the model parameters, and up to $4\times$ compression of the embedding table as well as activations. Among all tasks, we observed the highest performance loss for BERT fine-tuned on SQuAD. By probing into the Hessian based analysis as well as visualization, we show that this is related to the fact that current training/fine-tuning strategy of BERT does not converge for SQuAD.
In this paper, we propose a novel variable-rate learned image compression framework with a conditional autoencoder. Previous learning-based image compression methods mostly require training separate networks for different compression rates so they can yield compressed images of varying quality. In contrast, we train and deploy only one variable-rate image compression network implemented with a conditional autoencoder. We provide two rate control parameters, i.e., the Lagrange multiplier and the quantization bin size, which are given as conditioning variables to the network. Coarse rate adaptation to a target is performed by changing the Lagrange multiplier, while the rate can be further fine-tuned by adjusting the bin size used in quantizing the encoded representation. Our experimental results show that the proposed scheme provides a better rate-distortion trade-off than the traditional variable-rate image compression codecs such as JPEG2000 and BPG. Our model also shows comparable and sometimes better performance than the state-of-the-art learned image compression models that deploy multiple networks trained for varying rates.
Applications of Fully Convolutional Networks (FCN) in iris segmentation have shown promising advances. For mobile and embedded systems, a significant challenge is that the proposed FCN architectures are extremely computationally demanding. In this article, we propose a resource-efficient, end-to-end iris recognition flow, which consists of FCN-based segmentation, contour fitting, followed by Daugman normalization and encoding. To attain accurate and efficient FCN models, we propose a three-step SW/HW co-design methodology consisting of FCN architectural exploration, precision quantization, and hardware acceleration. In our exploration, we propose multiple FCN models, and in comparison to previous works, our best-performing model requires 50X less FLOPs per inference while achieving a new state-of-the-art segmentation accuracy. Next, we select the most efficient set of models and further reduce their computational complexity through weights and activations quantization using 8-bit dynamic fixed-point (DFP) format. Each model is then incorporated into an end-to-end flow for true recognition performance evaluation. A few of our end-to-end pipelines outperform the previous state-of-the-art on two datasets evaluated. Finally, we propose a novel DFP accelerator and fully demonstrate the SW/HW co-design realization of our flow on an embedded FPGA platform. In comparison with the embedded CPU, our hardware acceleration achieves up to 8.3X speedup for the overall pipeline while using less than 15% of the available FPGA resources. We also provide comparisons between the FPGA system and an embedded GPU showing different benefits and drawbacks for the two platforms.
We consider the problem of the recovery of a k-sparse vector from compressed linear measurements when data are corrupted by a quantization noise. When the number of measurements is not sufficiently large, different $k$-sparse solutions may be present in the feasible set, and the classical l1 approach may be unsuccessful. For this motivation, we propose a non-convex quadratic programming method, which exploits prior information on the magnitude of the non-zero parameters. This results in a more efficient support recovery. We provide sufficient conditions for successful recovery and numerical simulations to illustrate the practical feasibility of the proposed method.
We present a comprehensive study and evaluation of existing single image compression artifacts removal algorithms, using a new 4K resolution benchmark including diversified foreground objects and background scenes with rich structures, called Large-scale Ideal Ultra high definition 4K (LIU4K) benchmark. Compression artifacts removal, as a common post-processing technique, aims at alleviating undesirable artifacts such as blockiness, ringing, and banding caused by quantization and approximation in the compression process. In this work, a systematic listing of the reviewed methods is presented based on their basic models (handcrafted models and deep networks). The main contributions and novelties of these methods are highlighted, and the main development directions, including architectures, multi-domain sources, signal structures, and new targeted units, are summarized. Furthermore, based on a unified deep learning configuration (i.e. same training data, loss function, optimization algorithm, etc.), we evaluate recent deep learning-based methods based on diversified evaluation measures. The experimental results show the state-of-the-art performance comparison of existing methods based on both full-reference, non-reference and task-driven metrics. Our survey would give a comprehensive reference source for future research on single image compression artifacts removal and inspire new directions of the related fields.
This paper develops a communication-efficient algorithm to solve the stochastic optimization problem defined over a distributed network, aiming at reducing the burdensome communication in applications such as distributed machine learning.Different from the existing works based on quantization and sparsification, we introduce a communication-censoring technique to reduce the transmissions of variables, which leads to our communication-Censored distributed Stochastic Gradient Descent (CSGD) algorithm. Specifically, in CSGD, the latest mini-batch stochastic gradient at a worker will be transmitted to the server if and only if it is sufficiently informative. When the latest gradient is not available, the stale one will be reused at the server. To implement this communication-censoring strategy, the batch-size is increasing in order to alleviate the effect of stochastic gradient noise. Theoretically, CSGD enjoys the same order of convergence rate as that of SGD, but effectively reduces communication. Numerical experiments demonstrate the sizable communication saving of CSGD.
The minimum error entropy (MEE) criterion has been verified as a powerful approach for non-Gaussian signal processing and robust machine learning. However, the implementation of MEE on robust classification is rather a vacancy in the literature. The original MEE only focuses on minimizing the Renyi's quadratic entropy of the error probability distribution function (PDF), which could cause failure in noisy classification tasks. To this end, we analyze the optimal error distribution in the presence of outliers for those classifiers with continuous errors, and introduce a simple codebook to restrict MEE so that it drives the error PDF towards the desired case. Half-quadratic based optimization and convergence analysis of the new learning criterion, called restricted MEE (RMEE), are provided. Experimental results with logistic regression and extreme learning machine are presented to verify the desirable robustness of RMEE.
The increasing inclusion of Deep Learning (DL) models in safety-critical systems such as autonomous vehicles have led to the development of multiple model-based DL testing techniques. One common denominator of these testing techniques is the automated generation of test cases, e.g., new inputs transformed from the original training data with the aim to optimize some test adequacy criteria. So far, the effectiveness of these approaches has been hindered by their reliance on random fuzzing or transformations that do not always produce test cases with a good diversity. To overcome these limitations, we propose, DeepEvolution, a novel search-based approach for testing DL models that relies on metaheuristics to ensure a maximum diversity in generated test cases. We assess the effectiveness of DeepEvolution in testing computer-vision DL models and found that it significantly increases the neuronal coverage of generated test cases. Moreover, using DeepEvolution, we could successfully find several corner-case behaviors. Finally, DeepEvolution outperformed Tensorfuzz (a coverage-guided fuzzing tool developed at Google Brain) in detecting latent defects introduced during the quantization of the models. These results suggest that search-based approaches can help build effective testing tools for DL systems.
This paper proposes a new subspace learning method, named Quantized Fisher Discriminant Analysis (QFDA), which makes use of both machine learning and information theory. There is a lack of literature for combination of machine learning and information theory and this paper tries to tackle this gap. QFDA finds a subspace which discriminates the uniformly quantized images in the Discrete Cosine Transform (DCT) domain at least as well as discrimination of non-quantized images by Fisher Discriminant Analysis (FDA) while the images have been compressed. This helps the user to throw away the original images and keep the compressed images instead without noticeable loss of classification accuracy. We propose a cost function whose minimization can be interpreted as rate-distortion optimization in information theory. We also propose quantized Fisherfaces for facial analysis in QFDA. Our experiments on AT&T face dataset and Fashion MNIST dataset show the effectiveness of this subspace learning method.
Deep neural network (DNN) quantization converting floating-point (FP) data in the network to integers (INT) is an effective way to shrink the model size for memory saving and simplify the operations for compute acceleration. Recently, researches on DNN quantization develop from inference to training, laying a foundation for the online training on accelerators. However, existing schemes leaving batch normalization (BN) untouched during training are mostly incomplete quantization that still adopts high precision FP in some parts of the data paths. Currently, there is no solution that can use only low bit-width INT data during the whole training process of large-scale DNNs with acceptable accuracy. In this work, through decomposing all the computation steps in DNNs and fusing three special quantization functions to satisfy the different precision requirements, we propose a unified complete quantization framework termed as ``WAGEUBN'' to quantize DNNs involving all data paths including W (Weights), A (Activation), G (Gradient), E (Error), U (Update), and BN. Moreover, the Momentum optimizer is also quantized to realize a completely quantized framework. Experiments on ResNet18/34/50 models demonstrate that WAGEUBN can achieve competitive accuracy on the ImageNet dataset. For the first time, the study of quantization in large-scale DNNs is advanced to the full 8-bit INT level. In this way, all the operations in the training and inference can be bit-wise operations, pushing towards faster processing speed, decreased memory cost, and higher energy efficiency. Our throughout quantization framework has great potential for future efficient portable devices with online learning ability.
Compressing word embeddings is important for deploying NLP models in memory-constrained settings. However, understanding what makes compressed embeddings perform well on downstream tasks is challenging---existing measures of compression quality often fail to distinguish between embeddings that perform well and those that do not. We thus propose the eigenspace overlap score as a new measure. We relate the eigenspace overlap score to downstream performance by developing generalization bounds for the compressed embeddings in terms of this score, in the context of linear and logistic regression. We then show that we can lower bound the eigenspace overlap score for a simple uniform quantization compression method, helping to explain the strong empirical performance of this method. Finally, we show that by using the eigenspace overlap score as a selection criterion between embeddings drawn from a representative set we compressed, we can efficiently identify the better performing embedding with up to $2\times$ lower selection error rates than the next best measure of compression quality, and avoid the cost of training a model for each task of interest.
Current massive datasets demand light-weight access for analysis. Discrete hashing methods are thus beneficial because they map high-dimensional data to compact binary codes that are efficient to store and process, while preserving semantic similarity. To optimize powerful deep learning methods for image hashing, gradient-based methods are required. Binary codes, however, are discrete and thus have no continuous derivatives. Relaxing the problem by solving it in a continuous space and then quantizing the solution is not guaranteed to yield separable binary codes. The quantization needs to be included in the optimization. In this paper we push for quantization: We optimize maximum class separability in the binary space. We introduce a margin on distances between dissimilar image pairs as measured in the binary space. In addition to pair-wise distances, we draw inspiration from Fisher's Linear Discriminant Analysis (Fisher LDA) to maximize the binary distances between classes and at the same time minimize the binary distance of images within the same class. Experiments on CIFAR-10, NUS-WIDE and ImageNet100 demonstrate compact codes comparing favorably to the current state of the art.
The high computation and memory storage of large deep neural networks (DNNs) models pose intensive challenges to the conventional Von-Neumann architecture, incurring substantial data movements in the memory hierarchy. The memristor crossbar array has emerged as a promising solution to mitigate the challenges and enable low-power acceleration of DNNs. Memristor-based weight pruning and weight quantization have been seperately investigated and proven effectiveness in reducing area and power consumption compared to the original DNN model. However, there has been no systematic investigation of memristor-based neuromorphic computing (NC) systems considering both weight pruning and weight quantization. In this paper, we propose an unified and systematic memristor-based framework considering both structured weight pruning and weight quantization by incorporating alternating direction method of multipliers (ADMM) into DNNs training. We consider hardware constraints such as crossbar blocks pruning, conductance range, and mismatch between weight value and real devices, to achieve high accuracy and low power and small area footprint. Our framework is mainly integrated by three steps, i.e., memristor-based ADMM regularized optimization, masked mapping and retraining. Experimental results show that our proposed framework achieves 29.81X (20.88X) weight compression ratio, with 98.38% (96.96%) and 98.29% (97.47%) power and area reduction on VGG-16 (ResNet-18) network where only have 0.5% (0.76%) accuracy loss, compared to the original DNN models. We share our models at link http://bit.ly/2Jp5LHJ.