This paper presents a new algorithm for the lossy compression of scalar data defined on 2D or 3D regular grids, with topological control. Certain techniques allow users to control the pointwise error induced by the compression. However, in many scenarios it is desirable to control in a similar way the preservation of higher-level notions, such as topological features , in order to provide guarantees on the outcome of post-hoc data analyses. This paper presents the first compression technique for scalar data which supports a strictly controlled loss of topological features. It provides users with specific guarantees both on the preservation of the important features and on the size of the smaller features destroyed during compression. In particular, we present a simple compression strategy based on a topologically adaptive quantization of the range. Our algorithm provides strong guarantees on the bottleneck distance between persistence diagrams of the input and decompressed data, specifically those associated with extrema. A simple extension of our strategy additionally enables a control on the pointwise error. We also show how to combine our approach with state-of-the-art compressors, to further improve the geometrical reconstruction. Extensive experiments, for comparable compression rates, demonstrate the superiority of our algorithm in terms of the preservation of topological features. We show the utility of our approach by illustrating the compatibility between the output of post-hoc topological data analysis pipelines, executed on the input and decompressed data, for simulated or acquired data sets. We also provide a lightweight VTK-based C++ implementation of our approach for reproduction purposes.
Compression is a key step to deploy large neural networks on resource-constrained platforms. As a popular compression technique, quantization constrains the number of distinct weight values and thus reducing the number of bits required to represent and store each weight. In this paper, we study the representation power of quantized neural networks. First, we prove the universal approximability of quantized ReLU networks on a wide class of functions. Then we provide upper bounds on the number of weights and the memory size for a given approximation error bound and the bit-width of weights for function-independent and function-dependent structures. Our results reveal that, to attain an approximation error bound of $\epsilon$, the number of weights needed by a quantized network is no more than $\mathcal{O}\left(\log^5(1/\epsilon)\right)$ times that of an unquantized network. This overhead is of much lower order than the lower bound of the number of weights needed for the error bound, supporting the empirical success of various quantization techniques. To the best of our knowledge, this is the first in-depth study on the complexity bounds of quantized neural networks.
Deep neural networks have evolved remarkably over the past few years and they are currently the fundamental tools of many intelligent systems. At the same time, the computational complexity and resource consumption of these networks also continue to increase. This will pose a significant challenge to the deployment of such networks, especially in real-time applications or on resource-limited devices. Thus, network acceleration has become a hot topic within the deep learning community. As for hardware implementation of deep neural networks, a batch of accelerators based on FPGA/ASIC have been proposed in recent years. In this paper, we provide a comprehensive survey of recent advances in network acceleration, compression and accelerator design from both algorithm and hardware points of view. Specifically, we provide a thorough analysis of each of the following topics: network pruning, low-rank approximation, network quantization, teacher-student networks, compact network design and hardware accelerators. Finally, we will introduce and discuss a few possible future directions.
The aim of this paper is to endow the well-known family of hypercubic quantization hashing methods with theoretical guarantees. In hypercubic quantization, applying a suitable (random or learned) rotation after dimensionality reduction has been experimentally shown to improve the results accuracy in the nearest neighbors search problem. We prove in this paper that the use of these rotations is optimal under some mild assumptions: getting optimal binary sketches is equivalent to applying a rotation uniformizing the diagonal of the covariance matrix between data points. Moreover, for two closed points, the probability to have dissimilar binary sketches is upper bounded by a factor of the initial distance between the data points. Relaxing these assumptions, we obtain a general concentration result for random matrices. We also provide some experiments illustrating these theoretical points and compare a set of algorithms in both the batch and online settings.
Given an existing trained neural network, it is often desirable to learn new capabilities without hindering performance of those already learned. Existing approaches either learn sub-optimal solutions, require joint training, or incur a substantial increment in the number of parameters for each added domain, typically as many as the original network. We propose a method called \emph{Deep Adaptation Networks} (DAN) that constrains newly learned filters to be linear combinations of existing ones. DANs precisely preserve performance on the original domain, require a fraction (typically 13\%, dependent on network architecture) of the number of parameters compared to standard fine-tuning procedures and converge in less cycles of training to a comparable or better level of performance. When coupled with standard network quantization techniques, we further reduce the parameter cost to around 3\% of the original with negligible or no loss in accuracy. The learned architecture can be controlled to switch between various learned representations, enabling a single network to solve a task from multiple different domains. We conduct extensive experiments showing the effectiveness of our method on a range of image classification tasks and explore different aspects of its behavior.
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.
In recent years, deep neural network exhibits its powerful superiority on information discrimination in many computer vision applications. However, the capacity of deep neural network architecture is still a mystery to the researchers. Intuitively, larger capacity of neural network can always deposit more information to improve the discrimination ability of the model. But, the learnable parameter scale is not feasible to estimate the capacity of deep neural network. Due to the overfitting, directly increasing hidden nodes number and hidden layer number are already demonstrated not necessary to effectively increase the network discrimination ability. In this paper, we propose a novel measurement, named "total valid bits", to evaluate the capacity of deep neural networks for exploring how to quantitatively understand the deep learning and the insights behind its super performance. Specifically, our scheme to retrieve the total valid bits incorporates the skilled techniques in both training phase and inference phase. In the network training, we design decimal weight regularization and 8-bit forward quantization to obtain the integer-oriented network representations. Moreover, we develop adaptive-bitwidth and non-uniform quantization strategy in the inference phase to find the neural network capacity, total valid bits. By allowing zero bitwidth, our adaptive-bitwidth quantization can execute the model reduction and valid bits finding simultaneously. In our extensive experiments, we first demonstrate that our total valid bits is a good indicator of neural network capacity. We also analyze the impact on network capacity from the network architecture and advanced training skills, such as dropout and batch normalization.
For unsupervised data-dependent hashing, the two most important requirements are to preserve similarity in the low-dimensional feature space and to minimize the binary quantization loss. A well-established hashing approach is Iterative Quantization (ITQ), which addresses these two requirements in separate steps. In this paper, we revisit the ITQ approach and propose novel formulations and algorithms to the problem. Specifically, we propose a novel approach, named Simultaneous Compression and Quantization (SCQ), to jointly learn to compress (reduce dimensionality) and binarize input data in a single formulation under strict orthogonal constraint. With this approach, we introduce a loss function and its relaxed version, termed Orthonormal Encoder (OnE) and Orthogonal Encoder (OgE) respectively, which involve challenging binary and orthogonal constraints. We propose to attack the optimization using novel algorithms based on recent advances in cyclic coordinate descent approach. Comprehensive experiments on unsupervised image retrieval demonstrate that our proposed methods consistently outperform other state-of-the-art hashing methods. Notably, our proposed methods outperform recent deep neural networks and GAN based hashing in accuracy, while being very computationally-efficient.
The performance and efficiency of distributed machine learning (ML) depends significantly on how long it takes for nodes to exchange state changes. Overly-aggressive attempts to reduce communication often sacrifice final model accuracy and necessitate additional ML techniques to compensate for this loss, limiting their generality. Some attempts to reduce communication incur high computation overhead, which makes their performance benefits visible only over slow networks. We present 3LC, a lossy compression scheme for state change traffic that strikes balance between multiple goals: traffic reduction, accuracy, computation overhead, and generality. It combines three new techniques---3-value quantization with sparsity multiplication, quartic encoding, and zero-run encoding---to leverage strengths of quantization and sparsification techniques and avoid their drawbacks. It achieves a data compression ratio of up to 39--107X, almost the same test accuracy of trained models, and high compression speed. Distributed ML frameworks can employ 3LC without modifications to existing ML algorithms. Our experiments show that 3LC reduces wall-clock training time of ResNet-110--based image classifiers for CIFAR-10 on a 10-GPU cluster by up to 16--23X compared to TensorFlow's baseline design.
Retinal vessel information is helpful in retinal disease screening and diagnosis. Retinal vessel segmentation provides useful information about vessels and can be used by physicians during intraocular surgery and retinal diagnostic operations. Convolutional neural networks (CNNs) are powerful tools for classification and segmentation of medical images. Complexity of CNNs makes it difficult to implement them in portable devices such as binocular indirect ophthalmoscopes. In this paper a simplification approach is proposed for CNNs based on combination of quantization and pruning. Fully connected layers are quantized and convolutional layers are pruned to have a simple and efficient network structure. Experiments on images of the STARE dataset show that our simplified network is able to segment retinal vessels with acceptable accuracy and low complexity.
This paper explores the problem of learning transforms for image compression via autoencoders. Usually, the rate-distortion performances of image compression are tuned by varying the quantization step size. In the case of autoen-coders, this in principle would require learning one transform per rate-distortion point at a given quantization step size. Here, we show that comparable performances can be obtained with a unique learned transform. The different rate-distortion points are then reached by varying the quantization step size at test time. This approach saves a lot of training time.
Persistent homology (PH) is a rigorous mathematical theory that provides a robust descriptor of data in the form of persistence diagrams (PDs) which are 2D multisets of points. Their variable size makes them, however, difficult to combine with typical machine learning workflows. In this paper we introduce persistence codebooks, a novel expressive and discriminative fixed-size vectorized representation of PDs. To this end, we adapt bag-of-words (BoW), vectors of locally aggregated descriptors (VLAD) and Fischer vectors (FV) for the quantization of PDs. Persistence codebooks represent PDs in a convenient way for machine learning and statistical analysis and have a number of favorable practical and theoretical properties including 1-Wasserstein stability. We evaluate the presented representations on several heterogeneous datasets and show their (high) discriminative power. Our approach achieves state-of-the-art performance and beyond in much less time than alternative approaches.
Batch Normalization (BN) has been proven to be quite effective at accelerating and improving the training of deep neural networks (DNNs). However, BN brings additional computation, consumes more memory and generally slows down the training process by a large margin, which aggravates the training effort. Furthermore, the nonlinear square and root operations in BN also impede the low bit-width quantization techniques, which draws much attention in deep learning hardware community. In this work, we propose an L1-norm BN (L1BN) with only linear operations in both the forward and the backward propagations during training. L1BN is shown to be approximately equivalent to the original L2-norm BN (L2BN) by multiplying a scaling factor. Experiments on various convolutional neural networks (CNNs) and generative adversarial networks (GANs) reveal that L1BN maintains almost the same accuracies and convergence rates compared to L2BN but with higher computational efficiency. On FPGA platform, the proposed signum and absolute operations in L1BN can achieve 1.5$\times$ speedup and save 50\% power consumption, compared with the original costly square and root operations, respectively. This hardware-friendly normalization method not only surpasses L2BN in speed, but also simplify the hardware design of ASIC accelerators with higher energy efficiency. Last but not the least, L1BN promises a fully quantized training of DNNs, which is crucial to future adaptive terminal devices.
Quantization can be used to form new vectors/matrices with shared values close to the original. In recent years, the popularity of scalar quantization for value-sharing applications has been soaring as it has been found huge utilities in reducing the complexity of neural networks. Existing clustering-based quantization techniques, while being well-developed, have multiple drawbacks including the dependency of the random seed, empty or out-of-the-range clusters, and high time complexity for a large number of clusters. To overcome these problems, in this paper, the problem of scalar quantization is examined from a new perspective, namely sparse least square optimization. Specifically, inspired by the property of sparse least square regression, several quantization algorithms based on $l_1$ least square are proposed. In addition, similar schemes with $l_1 + l_2$ and $l_0$ regularization are proposed. Furthermore, to compute quantization results with a given amount of values/clusters, this paper designed an iterative method and a clustering-based method, and both of them are built on sparse least square. The paper shows that the latter method is mathematically equivalent to an improved version of k-means clustering-based quantization algorithm, although the two algorithms originated from different intuitions. The algorithms proposed were tested with three types of data and their computational performances, including information loss, time consumption, and the distribution of the values of the sparse vectors, were compared and analyzed. The paper offers a new perspective to probe the area of quantization, and the algorithms proposed can outperform existing methods especially under some bit-width reduction scenarios, when the required post-quantization resolution (number of values) is not significantly lower than the original number.
Deploying state-of-the-art CNNs requires power-hungry processors and off-chip memory. This precludes the implementation of CNNs in low-power embedded systems. Recent research shows CNNs sustain extreme quantization, binarizing their weights and intermediate feature maps, thereby saving 8-32\x memory and collapsing energy-intensive sum-of-products into XNOR-and-popcount operations. We present XNORBIN, an accelerator for binary CNNs with computation tightly coupled to memory for aggressive data reuse. Implemented in UMC 65nm technology XNORBIN achieves an energy efficiency of 95 TOp/s/W and an area efficiency of 2.0 TOp/s/MGE at 0.8 V.
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.
High dynamic range (HDR) imaging provides the capability of handling real world lighting as opposed to the traditional low dynamic range (LDR) which struggles to accurately represent images with higher dynamic range. However, most imaging content is still available only in LDR. This paper presents a method for generating HDR content from LDR content based on deep Convolutional Neural Networks (CNNs) termed ExpandNet. ExpandNet accepts LDR images as input and generates images with an expanded range in an end-to-end fashion. The model attempts to reconstruct missing information that was lost from the original signal due to quantization, clipping, tone mapping or gamma correction. The added information is reconstructed from learned features, as the network is trained in a supervised fashion using a dataset of HDR images. The approach is fully automatic and data driven; it does not require any heuristics or human expertise. ExpandNet uses a multiscale architecture which avoids the use of upsampling layers to improve image quality. The method performs well compared to expansion/inverse tone mapping operators quantitatively on multiple metrics, even for badly exposed inputs.
Low-precision computation is often used to lower the time and energy cost of machine learning, and recently hardware accelerators have been developed to support it. Still, it has been used primarily for inference - not training. Previous low-precision training algorithms suffered from a fundamental tradeoff: as the number of bits of precision is lowered, quantization noise is added to the model, which limits statistical accuracy. To address this issue, we describe a simple low-precision stochastic gradient descent variant called HALP. HALP converges at the same theoretical rate as full-precision algorithms despite the noise introduced by using low precision throughout execution. The key idea is to use SVRG to reduce gradient variance, and to combine this with a novel technique called bit centering to reduce quantization error. We show that on the CPU, HALP can run up to $4 \times$ faster than full-precision SVRG and can match its convergence trajectory. We implemented HALP in TensorQuant, and show that it exceeds the validation performance of plain low-precision SGD on two deep learning tasks.
In this paper, we define and study a new notion of stability for the $k$-means clustering scheme building upon the notion of quantization of a probability measure. We connect this notion of stability to a geometric feature of the underlying distribution of the data, named absolute margin condition, inspired by recent works on the subject.
Although representation learning methods developed within the framework of traditional neural networks are relatively mature, developing a spiking representation model remains a challenging problem. This paper proposes an event-based method to train a feedforward spiking neural network (SNN) layer for extracting visual features. The method introduces a novel spike-timing-dependent plasticity (STDP) learning rule and a threshold adjustment rule both derived from a vector quantization-like objective function subject to a sparsity constraint. The STDP rule is obtained by the gradient of a vector quantization criterion that is converted to spike-based, spatio-temporally local update rules in a spiking network of leaky, integrate-and-fire (LIF) neurons. Independence and sparsity of the model are achieved by the threshold adjustment rule and by a softmax function implementing inhibition in the representation layer consisting of WTA-thresholded spiking neurons. Together, these mechanisms implement a form of spike-based, competitive learning. Two sets of experiments are performed on the MNIST and natural image datasets. The results demonstrate a sparse spiking visual representation model with low reconstruction loss comparable with state-of-the-art visual coding approaches, yet our rule is local in both time and space, thus biologically plausible and hardware friendly.