Recent research implies that training and inference of deep neural networks (DNN) can be computed with low precision numerical representations of the training/test data, weights and gradients without a general loss in accuracy. The benefit of such compact representations is twofold: they allow a significant reduction of the communication bottleneck in distributed DNN training and faster neural network implementations on hardware accelerators like FPGAs. Several quantization methods have been proposed to map the original 32-bit floating point problem to low-bit representations. While most related publications validate the proposed approach on a single DNN topology, it appears to be evident, that the optimal choice of the quantization method and number of coding bits is topology dependent. To this end, there is no general theory available, which would allow users to derive the optimal quantization during the design of a DNN topology. In this paper, we present a quantization tool box for the TensorFlow framework. TensorQuant allows a transparent quantization simulation of existing DNN topologies during training and inference. TensorQuant supports generic quantization methods and allows experimental evaluation of the impact of the quantization on single layers as well as on the full topology. In a first series of experiments with TensorQuant, we show an analysis of fix-point quantizations of popular CNN topologies.
This paper focuses on an examination of an applicability of Recurrent Neural Network models for detecting anomalous behavior of the CERN superconducting magnets. In order to conduct the experiments, the authors designed and implemented an adaptive signal quantization algorithm and a custom GRU-based detector and developed a method for the detector parameters selection. Three different datasets were used for testing the detector. Two artificially generated datasets were used to assess the raw performance of the system whereas the 231 MB dataset composed of the signals acquired from HiLumi magnets was intended for real-life experiments and model training. Several different setups of the developed anomaly detection system were evaluated and compared with state-of-the-art OC-SVM reference model operating on the same data. The OC-SVM model was equipped with a rich set of feature extractors accounting for a range of the input signal properties. It was determined in the course of the experiments that the detector, along with its supporting design methodology, reaches F1 equal or very close to 1 for almost all test sets. Due to the profile of the data, the best_length setup of the detector turned out to perform the best among all five tested configuration schemes of the detection system. The quantization parameters have the biggest impact on the overall performance of the detector with the best values of input/output grid equal to 16 and 8, respectively. The proposed solution of the detection significantly outperformed OC-SVM-based detector in most of the cases, with much more stable performance across all the datasets.
Neural networks with low-precision weights and activations offer compelling efficiency advantages over their full-precision equivalents. The two most frequently discussed benefits of quantization are reduced memory consumption, and a faster forward pass when implemented with efficient bitwise operations. We propose a third benefit of very low-precision neural networks: improved robustness against some adversarial attacks, and in the worst case, performance that is on par with full-precision models. We focus on the very low-precision case where weights and activations are both quantized to $\pm$1, and note that stochastically quantizing weights in just one layer can sharply reduce the impact of iterative attacks. We observe that non-scaled binary neural networks exhibit a similar effect to the original defensive distillation procedure that led to gradient masking, and a false notion of security. We address this by conducting both black-box and white-box experiments with binary models that do not artificially mask gradients.
We propose DoReFa-Net, a method to train convolutional neural networks that have low bitwidth weights and activations using low bitwidth parameter gradients. In particular, during backward pass, parameter gradients are stochastically quantized to low bitwidth numbers before being propagated to convolutional layers. As convolutions during forward/backward passes can now operate on low bitwidth weights and activations/gradients respectively, DoReFa-Net can use bit convolution kernels to accelerate both training and inference. Moreover, as bit convolutions can be efficiently implemented on CPU, FPGA, ASIC and GPU, DoReFa-Net opens the way to accelerate training of low bitwidth neural network on these hardware. Our experiments on SVHN and ImageNet datasets prove that DoReFa-Net can achieve comparable prediction accuracy as 32-bit counterparts. For example, a DoReFa-Net derived from AlexNet that has 1-bit weights, 2-bit activations, can be trained from scratch using 6-bit gradients to get 46.1\% top-1 accuracy on ImageNet validation set. The DoReFa-Net AlexNet model is released publicly.
Deep neural networks (DNNs) continue to make significant advances, solving tasks from image classification to translation or reinforcement learning. One aspect of the field receiving considerable attention is efficiently executing deep models in resource-constrained environments, such as mobile or embedded devices. This paper focuses on this problem, and proposes two new compression methods, which jointly leverage weight quantization and distillation of larger teacher networks into smaller student networks. The first method we propose is called quantized distillation and leverages distillation during the training process, by incorporating distillation loss, expressed with respect to the teacher, into the training of a student network whose weights are quantized to a limited set of levels. The second method, differentiable quantization, optimizes the location of quantization points through stochastic gradient descent, to better fit the behavior of the teacher model. We validate both methods through experiments on convolutional and recurrent architectures. We show that quantized shallow students can reach similar accuracy levels to full-precision teacher models, while providing order of magnitude compression, and inference speedup that is linear in the depth reduction. In sum, our results enable DNNs for resource-constrained environments to leverage architecture and accuracy advances developed on more powerful devices.
Network quantization is an effective solution to compress deep neural networks for practical usage. Existing network quantization methods cannot sufficiently exploit the depth information to generate low-bit compressed network. In this paper, we propose two novel network quantization approaches, single-level network quantization (SLQ) for high-bit quantization and multi-level network quantization (MLQ) for extremely low-bit quantization (ternary).We are the first to consider the network quantization from both width and depth level. In the width level, parameters are divided into two parts: one for quantization and the other for re-training to eliminate the quantization loss. SLQ leverages the distribution of the parameters to improve the width level. In the depth level, we introduce incremental layer compensation to quantize layers iteratively which decreases the quantization loss in each iteration. The proposed approaches are validated with extensive experiments based on the state-of-the-art neural networks including AlexNet, VGG-16, GoogleNet and ResNet-18. Both SLQ and MLQ achieve impressive results.
This work presents a method for adapting a single, fixed deep neural network to multiple tasks without affecting performance on already learned tasks. By building upon ideas from network quantization and pruning, we learn binary masks that piggyback on an existing network, or are applied to unmodified weights of that network to provide good performance on a new task. These masks are learned in an end-to-end differentiable fashion, and incur a low overhead of 1 bit per network parameter, per task. Even though the underlying network is fixed, the ability to mask individual weights allows for the learning of a large number of filters. We show performance comparable to dedicated fine-tuned networks for a variety of classification tasks, including those with large domain shifts from the initial task (ImageNet), and a variety of network architectures. Unlike prior work, we do not suffer from catastrophic forgetting or competition between tasks, and our performance is agnostic to task ordering. Code available at https://github.com/arunmallya/piggyback.
State-of-the-art human pose estimation methods are based on heat map representation. In spite of the good performance, the representation has a few issues in nature, such as not differentiable and quantization error. This work shows that a simple integral operation relates and unifies the heat map representation and joint regression, thus avoiding the above issues. It is differentiable, efficient, and compatible with any heat map based methods. Its effectiveness is convincingly validated via comprehensive ablation experiments under various settings, specifically on 3D pose estimation, for the first time.
Binarization is an extreme network compression approach that provides large computational speedups along with energy and memory savings, albeit at significant accuracy costs. We investigate the question of where to binarize inputs at layer-level granularity and show that selectively binarizing the inputs to specific layers in the network could lead to significant improvements in accuracy while preserving most of the advantages of binarization. We analyze the binarization tradeoff using a metric that jointly models the input binarization-error and computational cost and introduce an efficient algorithm to select layers whose inputs are to be binarized. Practical guidelines based on insights obtained from applying the algorithm to a variety of models are discussed. Experiments on Imagenet dataset using AlexNet and ResNet-18 models show 3-4% improvements in accuracy over fully binarized networks with minimal impact on compression and computational speed. The improvements are even more substantial on sketch datasets like TU-Berlin, where we match state-of-the-art accuracy as well, getting over 8% increase in accuracies. We further show that our approach can be applied in tandem with other forms of compression that deal with individual layers or overall model compression (e.g., SqueezeNets). Unlike previous quantization approaches, we are able to binarize the weights in the last layers of a network, which often have a large number of parameters, resulting in significant improvement in accuracy over fully binarized models.
As neural networks grow deeper and wider, learning networks with hard-threshold activations is becoming increasingly important, both for network quantization, which can drastically reduce time and energy requirements, and for creating large integrated systems of deep networks, which may have non-differentiable components and must avoid vanishing and exploding gradients for effective learning. However, since gradient descent is not applicable to hard-threshold functions, it is not clear how to learn networks of them in a principled way. We address this problem by observing that setting targets for hard-threshold hidden units in order to minimize loss is a discrete optimization problem, and can be solved as such. The discrete optimization goal is to find a set of targets such that each unit, including the output, has a linearly separable problem to solve. Given these targets, the network decomposes into individual perceptrons, which can then be learned with standard convex approaches. Based on this, we develop a recursive mini-batch algorithm for learning deep hard-threshold networks that includes the popular but poorly justified straight-through estimator as a special case. Empirically, we show that our algorithm improves classification accuracy in a number of settings, including for AlexNet and ResNet-18 on ImageNet, when compared to the straight-through estimator.
Offside detection in soccer has emerged as one of the most important decisions with an average of 50 offside decisions every game. False detections and rash calls adversely affect game conditions and in many cases drastically change the outcome of the game. The human eye has finite precision and can only discern a limited amount of detail in a given instance. Current offside decisions are made manually by sideline referees and tend to remain controversial in many games. This calls for automated offside detection techniques in order to assist accurate refereeing. In this work, we have explicitly used computer vision and image processing techniques like Hough transform, color similarity (quantization), graph connected components, and vanishing point ideas to identify the probable offside regions. Keywords: Hough transform, connected components, KLT tracking, color similarity.
Similarity search approaches based on graph walks have recently attained outstanding speed-accuracy trade-offs, taking aside the memory requirements. In this paper, we revisit these approaches by considering, additionally, the memory constraint required to index billions of images on a single server. This leads us to propose a method based both on graph traversal and compact representations. We encode the indexed vectors using quantization and exploit the graph structure to refine the similarity estimation. In essence, our method takes the best of these two worlds: the search strategy is based on nested graphs, thereby providing high precision with a relatively small set of comparisons. At the same time it offers a significant memory compression. As a result, our approach outperforms the state of the art on operating points considering 64-128 bytes per vector, as demonstrated by our results on two billion-scale public benchmarks.
The performance of speaker-related systems usually degrades heavily in practical applications largely due to the presence of background noise. To improve the robustness of such systems in unknown noisy environments, this paper proposes a simple pre-processing method called Noise Invariant Frame Selection (NIFS). Based on several noisy constraints, it selects noise invariant frames from utterances to represent speakers. Experiments conducted on the TIMIT database showed that the NIFS can significantly improve the performance of Vector Quantization (VQ), Gaussian Mixture Model-Universal Background Model (GMM-UBM) and i-vector-based speaker verification systems in different unknown noisy environments with different SNRs, in comparison to their baselines. Meanwhile, the proposed NIFS-based speaker verification systems achieves similar performance when we change the constraints (hyper-parameters) or features, which indicates that it is robust and easy to reproduce. Since NIFS is designed as a general algorithm, it could be further applied to other similar tasks.
The huge size of deep networks hinders their use in small computing devices. In this paper, we consider compressing the network by weight quantization. We extend a recently proposed loss-aware weight binarization scheme to ternarization, with possibly different scaling parameters for the positive and negative weights, and m-bit (where m > 2) quantization. Experiments on feedforward and recurrent neural networks show that the proposed scheme outperforms state-of-the-art weight quantization algorithms, and is as accurate (or even more accurate) than the full-precision network.
User and item features of side information are crucial for accurate recommendation. However, the large number of feature dimensions, e.g., usually larger than 10^7, results in expensive storage and computational cost. This prohibits fast recommendation especially on mobile applications where the computational resource is very limited. In this paper, we develop a generic feature-based recommendation model, called Discrete Factorization Machine (DFM), for fast and accurate recommendation. DFM binarizes the real-valued model parameters (e.g., float32) of every feature embedding into binary codes (e.g., boolean), and thus supports efficient storage and fast user-item score computation. To avoid the severe quantization loss of the binarization, we propose a convergent updating rule that resolves the challenging discrete optimization of DFM. Through extensive experiments on two real-world datasets, we show that 1) DFM consistently outperforms state-of-the-art binarized recommendation models, and 2) DFM shows very competitive performance compared to its real-valued version (FM), demonstrating the minimized quantization loss. This work is accepted by IJCAI 2018.
Deep Neural Networks trained as image auto-encoders have recently emerged as a promising direction for advancing the state-of-the-art in image compression. The key challenge in learning such networks is twofold: To deal with quantization, and to control the trade-off between reconstruction error (distortion) and entropy (rate) of the latent image representation. In this paper, we focus on the latter challenge and propose a new technique to navigate the rate-distortion trade-off for an image compression auto-encoder. The main idea is to directly model the entropy of the latent representation by using a context model: A 3D-CNN which learns a conditional probability model of the latent distribution of the auto-encoder. During training, the auto-encoder makes use of the context model to estimate the entropy of its representation, and the context model is concurrently updated to learn the dependencies between the symbols in the latent representation. Our experiments show that this approach, when measured in MS-SSIM, yields a state-of-the-art image compression system based on a simple convolutional auto-encoder.
Recently, many studies have demonstrated deep neural network (DNN) classifiers can be fooled by the adversarial example, which is crafted via introducing some perturbations into an original sample. Accordingly, some powerful defense techniques were proposed. However, existing defense techniques often require modifying the target model or depend on the prior knowledge of attacks. In this paper, we propose a straightforward method for detecting adversarial image examples, which can be directly deployed into unmodified off-the-shelf DNN models. We consider the perturbation to images as a kind of noise and introduce two classic image processing techniques, scalar quantization and smoothing spatial filter, to reduce its effect. The image entropy is employed as a metric to implement an adaptive noise reduction for different kinds of images. Consequently, the adversarial example can be effectively detected by comparing the classification results of a given sample and its denoised version, without referring to any prior knowledge of attacks. More than 20,000 adversarial examples against some state-of-the-art DNN models are used to evaluate the proposed method, which are crafted with different attack techniques. The experiments show that our detection method can achieve a high overall F1 score of 96.39% and certainly raises the bar for defense-aware attacks.
Deep reinforcement learning, applied to vision-based problems like Atari games, maps pixels directly to actions; internally, the deep neural network bears the responsibility of both extracting useful information and making decisions based on it. By separating the image processing from decision-making, one could better understand the complexity of each task, as well as potentially find smaller policy representations that are easier for humans to understand and may generalize better. To this end, we propose a new method for learning policies and compact state representations separately but simultaneously for policy approximation in reinforcement learning. State representations are generated by an encoder based on two novel algorithms: Increasing Dictionary Vector Quantization makes the encoder capable of growing its dictionary size over time, to address new observations as they appear in an open-ended online-learning context; Direct Residuals Sparse Coding encodes observations by disregarding reconstruction error minimization, and aiming instead for highest information inclusion. The encoder autonomously selects observations online to train on, in order to maximize code sparsity. As the dictionary size increases, the encoder produces increasingly larger inputs for the neural network: this is addressed by a variation of the Exponential Natural Evolution Strategies algorithm which adapts its probability distribution dimensionality along the run. We test our system on a selection of Atari games using tiny neural networks of only 6 to 18 neurons (depending on the game's controls). These are still capable of achieving results comparable---and occasionally superior---to state-of-the-art techniques which use two orders of magnitude more neurons.
Discretizing multi-dimensional data distributions is a fundamental step of modern indexing methods. State-of-the-art techniques learn parameters of quantizers on training data for optimal performance, thus adapting quantizers to the data. In this work, we propose to reverse this paradigm and adapt the data to the quantizer: we train a neural net which last layer forms a fixed parameter-free quantizer, such as pre-defined points of a hyper-sphere. As a proxy objective, we design and train a neural network that favors uniformity in the spherical latent space, while preserving the neighborhood structure after the mapping. We propose a new regularizer derived from the Kozachenko--Leonenko differential entropy estimator to enforce uniformity and combine it with a locality-aware triplet loss. Experiments show that our end-to-end approach outperforms most learned quantization methods, and is competitive with the state of the art on widely adopted benchmarks. Furthermore, we show that training without the quantization step results in almost no difference in accuracy, but yields a generic catalyzer that can be applied with any subsequent quantizer.
Quantized Neural Networks (QNNs) are often used to improve network efficiency during the inference phase, i.e. after the network has been trained. Extensive research in the field suggests many different quantization schemes. Still, the number of bits required, as well as the best quantization scheme, are yet unknown. Our theoretical analysis suggests that most of the training process is robust to substantial precision reduction, and points to only a few specific operations that require higher precision. Armed with this knowledge, we quantize the model parameters, activations and layer gradients to 8-bit, leaving at a higher precision only the final step in the computation of the weight gradients. Additionally, as QNNs require batch-normalization to be trained at high precision, we introduce Range Batch-Normalization (BN) which has significantly higher tolerance to quantization noise and improved computational complexity. Our simulations show that Range BN is equivalent to the traditional batch norm if a precise scale adjustment, which can be approximated analytically, is applied. To the best of the authors' knowledge, this work is the first to quantize the weights, activations, as well as a substantial volume of the gradients stream, in all layers (including batch normalization) to 8-bit while showing state-of-the-art results over the ImageNet-1K dataset.