Recently, a considerable amount of works have been made to tackle the communication burden in federated learning (FL) (e.g., model quantization, data sparsification, and model compression). However, the existing methods, that boost the communication efficiency in FL, result in a considerable trade-off between communication efficiency and global convergence rate. We formulate an optimization problem for compression-aided FL, which captures the relationship between the distortion rate, number of participating IoT devices, and convergence rate. Following that, the objective function is to minimize the total transmission time for FL convergence. Because the problem is non-convex, we propose to decompose it into sub-problems. Based on the property of a FL model, we first determine the number of IoT devices participating in the FL process. Then, the communication between IoT devices and the server is optimized by efficiently allocating wireless resources based on a coalition game. Our theoretical analysis shows that, by actively controlling the number of participating IoT devices, we can avoid the training divergence of compression-aided FL while maintaining the communication efficiency.
Low-Power Edge-AI capabilities are essential for on-device extended reality (XR) applications to support the vision of Metaverse. In this work, we investigate two representative XR workloads: (i) Hand detection and (ii) Eye segmentation, for hardware design space exploration. For both applications, we train deep neural networks and analyze the impact of quantization and hardware specific bottlenecks. Through simulations, we evaluate a CPU and two systolic inference accelerator implementations. Next, we compare these hardware solutions with advanced technology nodes. The impact of integrating state-of-the-art emerging non-volatile memory technology (STT/SOT/VGSOT MRAM) into the XR-AI inference pipeline is evaluated. We found that significant energy benefits (>=24%) can be achieved for hand detection (IPS=10) and eye segmentation (IPS=0.1) by introducing non-volatile memory in the memory hierarchy for designs at 7nm node while meeting minimum IPS (inference per second). Moreover, we can realize substantial reduction in area (>=30%) owing to the small form factor of MRAM compared to traditional SRAM.
Data clipping is crucial in reducing noise in quantization operations and improving the achievable accuracy of quantization-aware training (QAT). Current practices rely on heuristics to set clipping threshold scalars and cannot be shown to be optimal. We propose Optimally Clipped Tensors And Vectors (OCTAV), a recursive algorithm to determine MSE-optimal clipping scalars. Derived from the fast Newton-Raphson method, OCTAV finds optimal clipping scalars on the fly, for every tensor, at every iteration of the QAT routine. Thus, the QAT algorithm is formulated with provably minimum quantization noise at each step. In addition, we reveal limitations in common gradient estimation techniques in QAT and propose magnitude-aware differentiation as a remedy to further improve accuracy. Experimentally, OCTAV-enabled QAT achieves state-of-the-art accuracy on multiple tasks. These include training-from-scratch and retraining ResNets and MobileNets on ImageNet, and Squad fine-tuning using BERT models, where OCTAV-enabled QAT consistently preserves accuracy at low precision (4-to-6-bits). Our results require no modifications to the baseline training recipe, except for the insertion of quantization operations where appropriate.
Adversarial training (AT) and its variants have spearheaded progress in improving neural network robustness to adversarial perturbations and common corruptions in the last few years. Algorithm design of AT and its variants are focused on training models at a specified perturbation strength $\epsilon$ and only using the feedback from the performance of that $\epsilon$-robust model to improve the algorithm. In this work, we focus on models, trained on a spectrum of $\epsilon$ values. We analyze three perspectives: model performance, intermediate feature precision and convolution filter sensitivity. In each, we identify alternative improvements to AT that otherwise wouldn't have been apparent at a single $\epsilon$. Specifically, we find that for a PGD attack at some strength $\delta$, there is an AT model at some slightly larger strength $\epsilon$, but no greater, that generalizes best to it. Hence, we propose overdesigning for robustness where we suggest training models at an $\epsilon$ just above $\delta$. Second, we observe (across various $\epsilon$ values) that robustness is highly sensitive to the precision of intermediate features and particularly those after the first and second layer. Thus, we propose adding a simple quantization to defenses that improves accuracy on seen and unseen adaptive attacks. Third, we analyze convolution filters of each layer of models at increasing $\epsilon$ and notice that those of the first and second layer may be solely responsible for amplifying input perturbations. We present our findings and demonstrate our techniques through experiments with ResNet and WideResNet models on the CIFAR-10 and CIFAR-10-C datasets.
Gradient compression is a popular technique for improving communication complexity of stochastic first-order methods in distributed training of machine learning models. However, the existing works consider only with-replacement sampling of stochastic gradients. In contrast, it is well-known in practice and recently confirmed in theory that stochastic methods based on without-replacement sampling, e.g., Random Reshuffling (RR) method, perform better than ones that sample the gradients with-replacement. In this work, we close this gap in the literature and provide the first analysis of methods with gradient compression and without-replacement sampling. We first develop a na\"ive combination of random reshuffling with gradient compression (Q-RR). Perhaps surprisingly, but the theoretical analysis of Q-RR does not show any benefits of using RR. Our extensive numerical experiments confirm this phenomenon. This happens due to the additional compression variance. To reveal the true advantages of RR in the distributed learning with compression, we propose a new method called DIANA-RR that reduces the compression variance and has provably better convergence rates than existing counterparts with with-replacement sampling of stochastic gradients. Next, to have a better fit to Federated Learning applications, we incorporate local computation, i.e., we propose and analyze the variants of Q-RR and DIANA-RR -- Q-NASTYA and DIANA-NASTYA that use local gradient steps and different local and global stepsizes. Finally, we conducted several numerical experiments to illustrate our theoretical results.
A logconcave likelihood is as important to proper statistical inference as a convex cost function is important to variational optimization. Quantization is often disregarded when writing likelihood models, ignoring the limitations of the physical detectors used to collect the data. These two facts call for the question: would including quantization in likelihood models preclude logconcavity? are the true data likelihoods logconcave? We provide a general proof that the same simple assumption that leads to logconcave continuous-data likelihoods also leads to logconcave quantized-data likelihoods, provided that convex quantization regions are used.
Quantized neural networks have drawn a lot of attention as they reduce the space and computational complexity during the inference. Moreover, there has been folklore that quantization acts as an implicit regularizer and thus can improve the generalizability of neural networks, yet no existing work formalizes this interesting folklore. In this paper, we take the binary weights in a neural network as random variables under stochastic rounding, and study the distribution propagation over different layers in the neural network. We propose a quasi neural network to approximate the distribution propagation, which is a neural network with continuous parameters and smooth activation function. We derive the neural tangent kernel (NTK) for this quasi neural network, and show that the eigenvalue of NTK decays at approximately exponential rate, which is comparable to that of Gaussian kernel with randomized scale. This in turn indicates that the Reproducing Kernel Hilbert Space (RKHS) of a binary weight neural network covers a strict subset of functions compared with the one with real value weights. We use experiments to verify that the quasi neural network we proposed can well approximate binary weight neural network. Furthermore, binary weight neural network gives a lower generalization gap compared with real value weight neural network, which is similar to the difference between Gaussian kernel and Laplace kernel.
Recently, more and more images are compressed and sent to the back-end devices for the machine analysis tasks~(\textit{e.g.,} object detection) instead of being purely watched by humans. However, most traditional or learned image codecs are designed to minimize the distortion of the human visual system without considering the increased demand from machine vision systems. In this work, we propose a preprocessing enhanced image compression method for machine vision tasks to address this challenge. Instead of relying on the learned image codecs for end-to-end optimization, our framework is built upon the traditional non-differential codecs, which means it is standard compatible and can be easily deployed in practical applications. Specifically, we propose a neural preprocessing module before the encoder to maintain the useful semantic information for the downstream tasks and suppress the irrelevant information for bitrate saving. Furthermore, our neural preprocessing module is quantization adaptive and can be used in different compression ratios. More importantly, to jointly optimize the preprocessing module with the downstream machine vision tasks, we introduce the proxy network for the traditional non-differential codecs in the back-propagation stage. We provide extensive experiments by evaluating our compression method for two representative downstream tasks with different backbone networks. Experimental results show our method achieves a better trade-off between the coding bitrate and the performance of the downstream machine vision tasks by saving about 20% bitrate.
This work introduces Directed-Evolution (DE) method for sparsification of neural networks, where the relevance of parameters to the network accuracy is directly assessed and the parameters that produce the least effect on accuracy when tentatively zeroed are indeed zeroed. DE method avoids a potentially combinatorial explosion of all possible candidate sets of parameters to be zeroed in large networks by mimicking evolution in the natural world. DE uses a distillation context [5]. In this context, the original network is the teacher and DE evolves the student neural network to the sparsification goal while maintaining minimal divergence between teacher and student. After the desired sparsification level is reached in each layer of the network by DE, a variety of quantization alternatives are used on the surviving parameters to find the lowest number of bits for their representation with acceptable loss of accuracy. A procedure to find optimal distribution of quantization levels in each sparsified layer is presented. Suitable final lossless encoding of the surviving quantized parameters is used for the final parameter representation. DE was used in sample of representative neural networks using MNIST, FashionMNIST and COCO data sets with progressive larger networks. An 80 classes YOLOv3 with more than 60 million parameters network trained on COCO dataset reached 90% sparsification and correctly identifies and segments all objects identified by the original network with more than 80% confidence using 4bit parameter quantization. Compression between 40x and 80x. It has not escaped the authors that techniques from different methods can be nested. Once the best parameter set for sparsification is identified in a cycle of DE, a decision on zeroing only a sub-set of those parameters can be made using a combination of criteria like parameter magnitude and Hessian approximations.
In this paper we present Hyper-Dimensional Reconfigurable Analytics at the Tactical Edge (HyDRATE) using low-SWaP embedded hardware that can perform real-time reconfiguration at the edge leveraging non-MAC (free of floating-point MultiplyACcumulate operations) deep neural nets (DNN) combined with hyperdimensional (HD) computing accelerators. We describe the algorithm, trained quantized model generation, and simulated performance of a feature extractor free of multiply-accumulates feeding a hyperdimensional logic-based classifier. Then we show how performance increases with the number of hyperdimensions. We describe the realized low-SWaP FPGA hardware and embedded software system compared to traditional DNNs and detail the implemented hardware accelerators. We discuss the measured system latency and power, noise robustness due to use of learnable quantization and HD computing, actual versus simulated system performance for a video activity classification task and demonstration of reconfiguration on this same dataset. We show that reconfigurability in the field is achieved by retraining only the feed-forward HD classifier without gradient descent backpropagation (gradient-free), using few-shot learning of new classes at the edge. Initial work performed used LRCN DNN and is currently extended to use Two-stream DNN with improved performance.
Federated learning alleviates the privacy risk in distributed learning by transmitting only the local model updates to the central server. However, it faces challenges including statistical heterogeneity of clients' datasets and resource constraints of client devices, which severely impact the training performance and user experience. Prior works have tackled these challenges by combining personalization with model compression schemes including quantization and pruning. However, the pruning is data-dependent and thus must be done on the client side which requires considerable computation cost. Moreover, the pruning normally trains a binary supermask $\in \{0, 1\}$ which significantly limits the model capacity yet with no computation benefit. Consequently, the training requires high computation cost and a long time to converge while the model performance does not pay off. In this work, we propose HideNseek which employs one-shot data-agnostic pruning at initialization to get a subnetwork based on weights' synaptic saliency. Each client then optimizes a sign supermask $\in \{-1, +1\}$ multiplied by the unpruned weights to allow faster convergence with the same compression rates as state-of-the-art. Empirical results from three datasets demonstrate that compared to state-of-the-art, HideNseek improves inferences accuracies by up to 40.6\% while reducing the communication cost and training time by up to 39.7\% and 46.8\% respectively.
In order to deploy deep models in a computationally efficient manner, model quantization approaches have been frequently used. In addition, as new hardware that supports mixed bitwidth arithmetic operations, recent research on mixed precision quantization (MPQ) begins to fully leverage the capacity of representation by searching optimized bitwidths for different layers and modules in a network. However, previous studies mainly search the MPQ strategy in a costly scheme using reinforcement learning, neural architecture search, etc., or simply utilize partial prior knowledge for bitwidth assignment, which might be biased and sub-optimal. In this work, we present a novel Stochastic Differentiable Quantization (SDQ) method that can automatically learn the MPQ strategy in a more flexible and globally-optimized space with smoother gradient approximation. Particularly, Differentiable Bitwidth Parameters (DBPs) are employed as the probability factors in stochastic quantization between adjacent bitwidth choices. After the optimal MPQ strategy is acquired, we further train our network with entropy-aware bin regularization and knowledge distillation. We extensively evaluate our method for several networks on different hardware (GPUs and FPGA) and datasets. SDQ outperforms all state-of-the-art mixed or single precision quantization with a lower bitwidth and is even better than the full-precision counterparts across various ResNet and MobileNet families, demonstrating the effectiveness and superiority of our method.
Federated learning has been proposed as a privacy-preserving machine learning framework that enables multiple clients to collaborate without sharing raw data. However, client privacy protection is not guaranteed by design in this framework. Prior work has shown that the gradient sharing strategies in federated learning can be vulnerable to data reconstruction attacks. In practice, though, clients may not transmit raw gradients considering the high communication cost or due to privacy enhancement requirements. Empirical studies have demonstrated that gradient obfuscation, including intentional obfuscation via gradient noise injection and unintentional obfuscation via gradient compression, can provide more privacy protection against reconstruction attacks. In this work, we present a new data reconstruction attack framework targeting the image classification task in federated learning. We show that commonly adopted gradient postprocessing procedures, such as gradient quantization, gradient sparsification, and gradient perturbation, may give a false sense of security in federated learning. Contrary to prior studies, we argue that privacy enhancement should not be treated as a byproduct of gradient compression. Additionally, we design a new method under the proposed framework to reconstruct the image at the semantic level. We quantify the semantic privacy leakage and compare with conventional based on image similarity scores. Our comparisons challenge the image data leakage evaluation schemes in the literature. The results emphasize the importance of revisiting and redesigning the privacy protection mechanisms for client data in existing federated learning algorithms.
This paper presents an analysis of the Low-Complexity Acoustic Scene Classification task in DCASE 2022 Challenge. The task was a continuation from the previous years, but the low-complexity requirements were changed to the following: the maximum number of allowed parameters, including the zero-valued ones, was 128 K, with parameters being represented using INT8 numerical format; and the maximum number of multiply-accumulate operations at inference time was 30 million. The provided baseline system is a convolutional neural network which employs post-training quantization of parameters, resulting in 46.5 K parameters, and 29.23 million multiply-and-accumulate operations (MMACs). Its performance on the evaluation data is 44.2% accuracy and 1.532 log-loss. In comparison, the top system in the challenge obtained an accuracy of 59.6% and a log loss of 1.091, having 121 K parameters and 28 MMACs. The task received 48 submissions from 19 different teams, most of which outperformed the baseline system.
Strong physical unclonable functions (PUFs) provide a low-cost authentication primitive for resource constrained devices. However, most strong PUF architectures can be modeled through learning algorithms with a limited number of CRPs. In this paper, we introduce the concept of non-monotonic response quantization for strong PUFs. Responses depend not only on which path is faster, but also on the distance between the arriving signals. Our experiments show that the resulting PUF has increased security against learning attacks. To demonstrate, we designed and implemented a non-monotonically quantized ring-oscillator based PUF in 65 nm technology. Measurement results show nearly ideal uniformity and uniqueness, with bit error rate of 13.4% over the temperature range from 0 C to 50 C.
Learning vectorized embeddings is at the core of various recommender systems for user-item matching. To perform efficient online inference, representation quantization, aiming to embed the latent features by a compact sequence of discrete numbers, recently shows the promising potentiality in optimizing both memory and computation overheads. However, existing work merely focuses on numerical quantization whilst ignoring the concomitant information loss issue, which, consequently, leads to conspicuous performance degradation. In this paper, we propose a novel quantization framework to learn Binarized Graph Representations for Top-K Recommendation (BiGeaR). BiGeaR introduces multi-faceted quantization reinforcement at the pre-, mid-, and post-stage of binarized representation learning, which substantially retains the representation informativeness against embedding binarization. In addition to saving the memory footprint, BiGeaR further develops solid online inference acceleration with bitwise operations, providing alternative flexibility for the realistic deployment. The empirical results over five large real-world benchmarks show that BiGeaR achieves about 22%~40% performance improvement over the state-of-the-art quantization-based recommender system, and recovers about 95%~102% of the performance capability of the best full-precision counterpart with over 8x time and space reduction.
How to efficiently serve ever-larger trained natural language models in practice has become exceptionally challenging even for powerful cloud servers due to their prohibitive memory/computation requirements. In this work, we present an efficient and affordable post-training quantization approach to compress large Transformer-based models, termed as ZeroQuant. ZeroQuant is an end-to-end quantization and inference pipeline with three main components: (1) a fine-grained hardware-friendly quantization scheme for both weight and activations; (2) a novel affordable layer-by-layer knowledge distillation algorithm (LKD) even without the access to the original training data; (3) a highly-optimized quantization system backend support to remove the quantization/dequantization overhead. As such, we are able to show that: (1) ZeroQuant can reduce the precision for weights and activations to INT8 in a cost-free way for both BERT and GPT3-style models with minimal accuracy impact, which leads to up to 5.19x/4.16x speedup on those models compared to FP16 inference; (2) ZeroQuant plus LKD affordably quantize the weights in the fully-connected module to INT4 along with INT8 weights in the attention module and INT8 activations, resulting in 3x memory footprint reduction compared to the FP16 model; (3) ZeroQuant can be directly applied to two of the largest open-sourced language models, including GPT-J6B and GPT-NeoX20, for which our INT8 model achieves similar accuracy as the FP16 model but achieves up to 5.2x better efficiency.
Extreme compression, particularly ultra-low bit precision (binary/ternary) quantization, has been proposed to fit large NLP models on resource-constraint devices. However, to preserve the accuracy for such aggressive compression schemes, cutting-edge methods usually introduce complicated compression pipelines, e.g., multi-stage expensive knowledge distillation with extensive hyperparameter tuning. Also, they oftentimes focus less on smaller transformer models that have already been heavily compressed via knowledge distillation and lack a systematic study to show the effectiveness of their methods. In this paper, we perform a very comprehensive systematic study to measure the impact of many key hyperparameters and training strategies from previous works. As a result, we find out that previous baselines for ultra-low bit precision quantization are significantly under-trained. Based on our study, we propose a simple yet effective compression pipeline for extreme compression, named XTC. XTC demonstrates that (1) we can skip the pre-training knowledge distillation to obtain a 5-layer BERT while achieving better performance than previous state-of-the-art methods, e.g., the 6-layer TinyBERT; (2) extreme quantization plus layer reduction is able to reduce the model size by 50x, resulting in new state-of-the-art results on GLUE tasks.
The success of deep learning comes from its ability to capture the hierarchical structure of data by learning high-level representations defined in terms of low-level ones. In this paper we explore self-supervised learning of hierarchical representations of speech by applying multiple levels of Contrastive Predictive Coding (CPC). We observe that simply stacking two CPC models does not yield significant improvements over single-level architectures. Inspired by the fact that speech is often described as a sequence of discrete units unevenly distributed in time, we propose a model in which the output of a low-level CPC module is non-uniformly downsampled to directly minimize the loss of a high-level CPC module. The latter is designed to also enforce a prior of separability and discreteness in its representations by enforcing dissimilarity of successive high-level representations through focused negative sampling, and by quantization of the prediction targets. Accounting for the structure of the speech signal improves upon single-level CPC features and enhances the disentanglement of the learned representations, as measured by downstream speech recognition tasks, while resulting in a meaningful segmentation of the signal that closely resembles phone boundaries.
This work studies federated learning (FL) over a fog radio access network, in which multiple internet-of-things (IoT) devices cooperatively learn a shared machine learning model by communicating with a cloud server (CS) through distributed access points (APs). Under the assumption that the fronthaul links connecting APs to CS have finite capacity, a rate-splitting transmission at IoT devices (IDs) is proposed which enables hybrid edge and cloud decoding of split uplink messages. The problem of completion time minimization for FL is tackled by optimizing the rate-splitting transmission and fronthaul quantization strategies along with training hyperparameters such as precision and iteration numbers. Numerical results show that the proposed rate-splitting transmission achieves notable gains over benchmark schemes which rely solely on edge or cloud decoding.