Network pruning and quantization are proven to be effective ways for deep model compression. To obtain a highly compact model, most methods first perform network pruning and then conduct network quantization based on the pruned model. However, this strategy may ignore that they would affect each other and thus performing them separately may lead to sub-optimal performance. To address this, performing pruning and quantization jointly is essential. Nevertheless, how to make a trade-off between pruning and quantization is non-trivial. Moreover, existing compression methods often rely on some pre-defined compression configurations. Some attempts have been made to search for optimal configurations, which however may take unbearable optimization cost. To address the above issues, we devise a simple yet effective method named Single-path Bit Sharing (SBS). Specifically, we first consider network pruning as a special case of quantization, which provides a unified view for pruning and quantization. We then introduce a single-path model to encode all candidate compression configurations. In this way, the configuration search problem is transformed into a subset selection problem, which significantly reduces the number of parameters, computational cost and optimization difficulty. Relying on the single-path model, we further introduce learnable binary gates to encode the choice of bitwidth. By jointly training the binary gates in conjunction with network parameters, the compression configurations of each layer can be automatically determined. Extensive experiments on both CIFAR-100 and ImageNet show that SBS is able to significantly reduce computational cost while achieving promising performance. For example, our SBS compressed MobileNetV2 achieves 22.6x Bit-Operation (BOP) reduction with only 0.1% drop in the Top-1 accuracy.
The continuous convergence of machine learning algorithms, 5G and beyond (5G+) wireless communications, and artificial intelligence (AI) hardware implementation hastens the birth of federated learning (FL) over 5G+ mobile devices, which pushes AI functions to mobile devices and initiates a new era of on-device AI applications. Despite the remarkable progress made in FL, huge energy consumption is one of the most significant obstacles restricting the development of FL over battery-constrained 5G+ mobile devices. To address this issue, in this paper, we investigate how to develop energy efficient FL over 5G+ mobile devices by making a trade-off between energy consumption for "working" (i.e., local computing) and that for "talking" (i.e., wireless communications) in order to boost the overall energy efficiency. Specifically, we first examine energy consumption models for graphics processing unit (GPU) computation and wireless transmissions. Then, we overview the state of the art of integrating FL procedure with energy-efficient learning techniques (e.g., gradient sparsification, weight quantization, pruning, etc.). Finally, we present several potential future research directions for FL over 5G+ mobile devices from the perspective of energy efficiency.
In this work, we present an energy-efficient distributed learning framework using low-resolution ADCs and coarsely quantized signals for Internet of Things (IoT) networks. In particular, we develop a distributed quantization-aware least-mean square (DQA-LMS) algorithm that can learn parameters in an energy-efficient fashion using signals quantized with few bits while requiring a low computational cost. We also carry out a statistical analysis of the proposed DQA-LMS algorithm that includes a stability condition. Simulations assess the DQA-LMS algorithm against existing techniques for a distributed parameter estimation task where IoT devices operate in a peer-to-peer mode and demonstrate the effectiveness of the DQA-LMS algorithm.
Time-to-contact (TTC), the time for an object to collide with the observer's plane, is a powerful tool for path planning: it is potentially more informative than the depth, velocity, and acceleration of objects in the scene -- even for humans. TTC presents several advantages, including requiring only a monocular, uncalibrated camera. However, regressing TTC for each pixel is not straightforward, and most existing methods make over-simplifying assumptions about the scene. We address this challenge by estimating TTC via a series of simpler, binary classifications. We predict with low latency whether the observer will collide with an obstacle within a certain time, which is often more critical than knowing exact, per-pixel TTC. For such scenarios, our method offers a temporal geofence in 6.4 ms -- over 25x faster than existing methods. Our approach can also estimate per-pixel TTC with arbitrarily fine quantization (including continuous values), when the computational budget allows for it. To the best of our knowledge, our method is the first to offer TTC information (binary or coarsely quantized) at sufficiently high frame-rates for practical use.
As neural networks gain widespread adoption in embedded devices, there is a need for model compression techniques to facilitate deployment in resource-constrained environments. Quantization is one of the go-to methods yielding state-of-the-art model compression. Most approaches take a fully trained model, apply different heuristics to determine the optimal bit-precision for different layers of the network, and retrain the network to regain any drop in accuracy. Based on Activation Density (AD)-the proportion of non-zero activations in a layer-we propose an in-training quantization method. Our method calculates bit-width for each layer during training yielding a mixed precision model with competitive accuracy. Since we train lower precision models during training, our approach yields the final quantized model at lower training complexity and also eliminates the need for re-training. We run experiments on benchmark datasets like CIFAR-10, CIFAR-100, TinyImagenet on VGG19/ResNet18 architectures and report the accuracy and energy estimates for the same. We achieve ~4.5x benefit in terms of estimated multiply-and-accumulate (MAC) reduction while reducing the training complexity by 50% in our experiments. To further evaluate the energy benefits of our proposed method, we develop a mixed-precision scalable Process In Memory (PIM) hardware accelerator platform. The hardware platform incorporates shift-add functionality for handling multi-bit precision neural network models. Evaluating the quantized models obtained with our proposed method on the PIM platform yields ~5x energy reduction compared to 16-bit models. Additionally, we find that integrating AD based quantization with AD based pruning (both conducted during training) yields up to ~198x and ~44x energy reductions for VGG19 and ResNet18 architectures respectively on PIM platform compared to baseline 16-bit precision, unpruned models.
Sound event detection (SED) is a hot topic in consumer and smart city applications. Existing approaches based on Deep Neural Networks are very effective, but highly demanding in terms of memory, power, and throughput when targeting ultra-low power always-on devices. Latency, availability, cost, and privacy requirements are pushing recent IoT systems to process the data on the node, close to the sensor, with a very limited energy supply, and tight constraints on the memory size and processing capabilities precluding to run state-of-the-art DNNs. In this paper, we explore the combination of extreme quantization to a small-footprint binary neural network (BNN) with the highly energy-efficient, RISC-V-based (8+1)-core GAP8 microcontroller. Starting from an existing CNN for SED whose footprint (815 kB) exceeds the 512 kB of memory available on our platform, we retrain the network using binary filters and activations to match these memory constraints. (Fully) binary neural networks come with a natural drop in accuracy of 12-18% on the challenging ImageNet object recognition challenge compared to their equivalent full-precision baselines. This BNN reaches a 77.9% accuracy, just 7% lower than the full-precision version, with 58 kB (7.2 times less) for the weights and 262 kB (2.4 times less) memory in total. With our BNN implementation, we reach a peak throughput of 4.6 GMAC/s and 1.5 GMAC/s over the full network, including preprocessing with Mel bins, which corresponds to an efficiency of 67.1 GMAC/s/W and 31.3 GMAC/s/W, respectively. Compared to the performance of an ARM Cortex-M4 implementation, our system has a 10.3 times faster execution time and a 51.1 times higher energy-efficiency.
In this paper we outline the development methodology for an automatic dog treat dispenser which combines machine learning and embedded hardware to identify and reward dog behaviors in real-time. Using machine learning techniques for training an image classification model we identify three behaviors of our canine companions: "sit", "stand", and "lie down" with up to 92% test accuracy and 39 frames per second. We evaluate a variety of neural network architectures, interpretability methods, model quantization and optimization techniques to develop a model specifically for an NVIDIA Jetson Nano. We detect the aforementioned behaviors in real-time and reinforce positive actions by making inference on the Jetson Nano and transmitting a signal to a servo motor to release rewards from a treat delivery apparatus.
Neural networks have achieved remarkable performance in computer vision, however they are vulnerable to adversarial examples. Adversarial examples are inputs that have been carefully perturbed to fool classifier networks, while appearing unchanged to humans. Based on prior works on detecting adversaries, we propose a structured methodology of augmenting a deep neural network (DNN) with a detector subnetwork. We use $\textit{Adversarial Noise Sensitivity}$ (ANS), a novel metric for measuring the adversarial gradient contribution of different intermediate layers of a network. Based on the ANS value, we append a detector to the most sensitive layer. In prior works, more complex detectors were added to a DNN, increasing the inference computational cost of the model. In contrast, our structured and strategic addition of a detector to a DNN reduces the complexity of the model while making the overall network adversarially resilient. Through comprehensive white-box and black-box experiments on MNIST, CIFAR-10, and CIFAR-100, we show that our method improves state-of-the-art detector robustness against adversarial examples. Furthermore, we validate the energy efficiency of our proposed adversarial detection methodology through an extensive energy analysis on various hardware scalable CMOS accelerator platforms. We also demonstrate the effects of quantization on our detector-appended networks.
Transformer based models, like BERT and RoBERTa, have achieved state-of-the-art results in many Natural Language Processing tasks. However, their memory footprint, inference latency, and power consumption are prohibitive efficient inference at the edge, and even at the data center. While quantization can be a viable solution for this, previous work on quantizing Transformer based models use floating-point arithmetic during inference, which cannot efficiently utilize integer-only logical units such as the recent Turing Tensor Cores, or traditional integer-only ARM processors. In this work, we propose I-BERT, a novel quantization scheme for Transformer based models that quantizes the entire inference with integer-only arithmetic. Based on lightweight integer-only approximation methods for nonlinear operations, e.g., GELU, Softmax, and Layer Normalization, I-BERT performs an end-to-end integer-only BERT inference without any floating point calculation. We evaluate our approach on GLUE downstream tasks using RoBERTa-Base/Large. We show that for both cases, I-BERT achieves similar (and slightly higher) accuracy as compared to the full-precision baseline. Furthermore, our preliminary implementation of I-BERT shows a speedup of 2.4-4.0x for INT8 inference on a T4 GPU system as compared to FP32 inference. The framework has been developed in PyTorch and has been open-sourced.
Ensembles of thermostatically controlled loads (TCL) provide a significant demand response reserve for the system operator to balance power grids. However, this also results in the parasitic synchronization of individual devices within the ensemble leading to long post-demand-response oscillations in the integrated energy consumption of the ensemble. The synchronization is eventually destructed by fluctuations, thus leading to the (pre-demand response) steady state; however, this natural desynchronization, or relaxation to a statistically steady state, is too long. A resolution of this problem consists in measuring the ensemble's instantaneous consumption and using it as a feedback to stochastic switching of the ensemble's devices between on- and off- states. A simplified continuous-time model showed that carefully tuned nonlinear feedback results in a fast (super-) relaxation of the ensemble energy consumption. Since both state information and control signals are discrete, the actual TCL devices operation is space-time quantized, and this must be considered for realistic TCL ensemble modelling. Here, assuming that states are characterized by indoor temperature (quantifying comfort) and air conditioner regime (on, off), we construct a discrete model based on the probabilistic description of state transitions. We demonstrate that super-relaxation holds in such a more realistic setting, and that while it is stable against randomness in the stochastic matrix of the quantized model, it remains sensitive to the time discretization scheme. Aiming to achieve a balance between super-relaxation and customer's comfort, we analyze the dependence of super-relaxation on details of the space-time quantization, and provide a simple analytical criterion to avoid undesirable oscillations in consumption.
Recently deep learning-based image compression has shown the potential to outperform traditional codecs. However, most existing methods train multiple networks for multiple bit rates, which increase the implementation complexity. In this paper, we propose a new variable-rate image compression framework, which employs generalized octave convolutions (GoConv) and generalized octave transposed-convolutions (GoTConv) with built-in generalized divisive normalization (GDN) and inverse GDN (IGDN) layers. Novel GoConv- and GoTConv-based residual blocks are also developed in the encoder and decoder networks. Our scheme also uses a stochastic rounding-based scalar quantization. To further improve the performance, we encode the residual between the input and the reconstructed image from the decoder network as an enhancement layer. To enable a single model to operate with different bit rates and to learn multi-rate image features, a new objective function is introduced. Experimental results show that the proposed framework trained with variable-rate objective function outperforms the standard codecs such as H.265/HEVC-based BPG and state-of-the-art learning-based variable-rate methods.
Neural networks are getting deeper and more computation-intensive nowadays. Quantization is a useful technique in deploying neural networks on hardware platforms and saving computation costs with negligible performance loss. However, recent research reveals that neural network models, no matter full-precision or quantized, are vulnerable to adversarial attacks. In this work, we analyze both adversarial and quantization losses and then introduce criteria to evaluate them. We propose a boundary-based retraining method to mitigate adversarial and quantization losses together and adopt a nonlinear mapping method to defend against white-box gradient-based adversarial attacks. The evaluations demonstrate that our method can better restore accuracy after quantization than other baseline methods on both black-box and white-box adversarial attacks. The results also show that adversarial training suffers quantization loss and does not cooperate well with other training methods.
The rapid development of large pre-trained language models has greatly increased the demand for model compression techniques, among which quantization is a popular solution. In this paper, we propose BinaryBERT, which pushes BERT quantization to the limit by weight binarization. We find that a binary BERT is hard to be trained directly than a ternary counterpart due to its complex and irregular loss landscape. Therefore, we propose ternary weight splitting, which initializes BinaryBERT by equivalently splitting from a half-sized ternary network. The binary model thus inherits the good performance of the ternary one, and can be further enhanced by fine-tuning the new architecture after splitting. Empirical results show that our BinaryBERT has only a slight performance drop compared with the full-precision model while being 24x smaller, achieving the state-of-the-art compression results on the GLUE and SQuAD benchmarks.
Recently, retrieval systems based on dense representations have led to important improvements in open-domain question answering, and related tasks. While very effective, this approach is also memory intensive, as the dense vectors for the whole knowledge source need to be kept in memory. In this paper, we study how the memory footprint of dense retriever-reader systems can be reduced. We consider three strategies to reduce the index size: dimension reduction, vector quantization and passage filtering. We evaluate our approach on two question answering benchmarks: TriviaQA and NaturalQuestions, showing that it is possible to get competitive systems using less than 6Gb of memory.
Unsupervised cross-modal hashing (UCMH) has become a hot topic recently. Current UCMH focuses on exploring data similarities. However, current UCMH methods calculate the similarity between two data, mainly relying on the two data's cross-modal features. These methods suffer from inaccurate similarity problems that result in a suboptimal retrieval Hamming space, because the cross-modal features between the data are not sufficient to describe the complex data relationships, such as situations where two data have different feature representations but share the inherent concepts. In this paper, we devise a deep graph-neighbor coherence preserving network (DGCPN). Specifically, DGCPN stems from graph models and explores graph-neighbor coherence by consolidating the information between data and their neighbors. DGCPN regulates comprehensive similarity preserving losses by exploiting three types of data similarities (i.e., the graph-neighbor coherence, the coexistent similarity, and the intra- and inter-modality consistency) and designs a half-real and half-binary optimization strategy to reduce the quantization errors during hashing. Essentially, DGCPN addresses the inaccurate similarity problem by exploring and exploiting the data's intrinsic relationships in a graph. We conduct extensive experiments on three public UCMH datasets. The experimental results demonstrate the superiority of DGCPN, e.g., by improving the mean average precision from 0.722 to 0.751 on MIRFlickr-25K using 64-bit hashing codes to retrieve texts from images. We will release the source code package and the trained model on https://github.com/Atmegal/DGCPN.
This paper proposes two novel techniques to train deep convolutional neural networks with low bit-width weights and activations. First, to obtain low bit-width weights, most existing methods obtain the quantized weights by performing quantization on the full-precision network weights. However, this approach would result in some mismatch: the gradient descent updates full-precision weights, but it does not update the quantized weights. To address this issue, we propose a novel method that enables {direct} updating of quantized weights {with learnable quantization levels} to minimize the cost function using gradient descent. Second, to obtain low bit-width activations, existing works consider all channels equally. However, the activation quantizers could be biased toward a few channels with high-variance. To address this issue, we propose a method to take into account the quantization errors of individual channels. With this approach, we can learn activation quantizers that minimize the quantization errors in the majority of channels. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on the image classification task, using AlexNet, ResNet and MobileNetV2 architectures on CIFAR-100 and ImageNet datasets.
Existing quantization aware training methods attempt to compensate for the quantization loss by leveraging on training data, like most of the post-training quantization methods, and are also time consuming. Both these methods are not effective for privacy constraint applications as they are tightly coupled with training data. In contrast, this paper proposes a data-independent post-training quantization scheme that eliminates the need for training data. This is achieved by generating a faux dataset, hereafter referred to as Retro-Synthesis Data, from the FP32 model layer statistics and further using it for quantization. This approach outperformed state-of-the-art methods including, but not limited to, ZeroQ and DFQ on models with and without Batch-Normalization layers for 8, 6, and 4 bit precisions on ImageNet and CIFAR-10 datasets. We also introduced two futuristic variants of post-training quantization methods namely Hybrid Quantization and Non-Uniform Quantization
Recent breakthroughs in deep neural networks (DNNs) have fueled a tremendous demand for intelligent edge devices featuring on-site learning, while the practical realization of such systems remains a challenge due to the limited resources available at the edge and the required massive training costs for state-of-the-art (SOTA) DNNs. As reducing precision is one of the most effective knobs for boosting training time/energy efficiency, there has been a growing interest in low-precision DNN training. In this paper, we explore from an orthogonal direction: how to fractionally squeeze out more training cost savings from the most redundant bit level, progressively along the training trajectory and dynamically per input. Specifically, we propose FracTrain that integrates (i) progressive fractional quantization which gradually increases the precision of activations, weights, and gradients that will not reach the precision of SOTA static quantized DNN training until the final training stage, and (ii) dynamic fractional quantization which assigns precisions to both the activations and gradients of each layer in an input-adaptive manner, for only "fractionally" updating layer parameters. Extensive simulations and ablation studies (six models, four datasets, and three training settings including standard, adaptation, and fine-tuning) validate the effectiveness of FracTrain in reducing computational cost and hardware-quantified energy/latency of DNN training while achieving a comparable or better (-0.12%~+1.87%) accuracy. For example, when training ResNet-74 on CIFAR-10, FracTrain achieves 77.6% and 53.5% computational cost and training latency savings, respectively, compared with the best SOTA baseline, while achieving a comparable (-0.07%) accuracy. Our codes are available at: https://github.com/RICE-EIC/FracTrain.
In this work, we introduce EQ-Net: the first holistic framework that solves both the tasks of log-likelihood ratio (LLR) estimation and quantization using a data-driven method. We motivate our approach with theoretical insights on two practical estimation algorithms at the ends of the complexity spectrum and reveal a connection between the complexity of an algorithm and the information bottleneck method: simpler algorithms admit smaller bottlenecks when representing their solution. This motivates us to propose a two-stage algorithm that uses LLR compression as a pretext task for estimation and is focused on low-latency, high-performance implementations via deep neural networks. We carry out extensive experimental evaluation and demonstrate that our single architecture achieves state-of-the-art results on both tasks when compared to previous methods, with gains in quantization efficiency as high as $20\%$ and reduced estimation latency by up to $60\%$ when measured on general purpose and graphical processing units (GPU). In particular, our approach reduces the GPU inference latency by more than two times in several multiple-input multiple-output (MIMO) configurations. Finally, we demonstrate that our scheme is robust to distributional shifts and retains a significant part of its performance when evaluated on 5G channel models, as well as channel estimation errors.
Quantizing deep convolutional neural networks for image super-resolution substantially reduces their computational costs. However, existing works either suffer from a severe performance drop in ultra-low precision of 4 or lower bit-widths, or require a heavy fine-tuning process to recover the performance. To our knowledge, this vulnerability to low precisions relies on two statistical observations of feature map values. First, distribution of feature map values varies significantly per channel and per input image. Second, feature maps have outliers that can dominate the quantization error. Based on these observations, we propose a novel distribution-aware quantization scheme (DAQ) which facilitates accurate training-free quantization in ultra-low precision. A simple function of DAQ determines dynamic range of feature maps and weights with low computational burden. Furthermore, our method enables mixed-precision quantization by calculating the relative sensitivity of each channel, without any training process involved. Nonetheless, quantization-aware training is also applicable for auxiliary performance gain. Our new method outperforms recent training-free and even training-based quantization methods to the state-of-the-art image super-resolution networks in ultra-low precision.