Federated Learning marks a turning point in the implementation of decentralized machine learning (especially deep learning) for wireless devices by protecting users' privacy and safeguarding raw data from third-party access. It assigns the learning process independently to each client. First, clients locally train a machine learning model based on local data. Next, clients transfer local updates of model weights and biases (training data) to a server. Then, the server aggregates updates (received from clients) to create a global learning model. However, the continuous transfer between clients and the server increases communication costs and is inefficient from a resource utilization perspective due to the large number of parameters (weights and biases) used by deep learning models. The cost of communication becomes a greater concern when the number of contributing clients and communication rounds increases. In this work, we propose a novel framework, FedZip, that significantly decreases the size of updates while transferring weights from the deep learning model between clients and their servers. FedZip implements Top-z sparsification, uses quantization with clustering, and implements compression with three different encoding methods. FedZip outperforms state-of-the-art compression frameworks and reaches compression rates up to 1085x, and preserves up to 99% of bandwidth and 99% of energy for clients during communication.
Machine Learning compilers like TVM allow a fast and flexible deployment on embedded CPUs. This enables the use of non-standard operators, which are common in ML compression techniques. However, it is necessary to understand the limitations of typical compute-intense operators in ML workloads to design a proper solution. This is the first in-detail analysis of dense and convolution operators, generated with TVM, that compares to the fundamental hardware limits of embedded ARM processors. Thereby it explains the gap between computational peak performance, theoretical and measured, and real-world state-of-the-art results, created with TVM and openBLAS. Instead, one can see that single-precision general matrix multiply (GEMM) and convolutions are bound by L1-cache-read bandwidth. Explorations of 8-bit and bit-serial quantized operators show that quantization can be used to achieve relevant speedups compared to cache-bound floating-point operators. However, the performance of quantized operators highly depends on the interaction between data layout and bit packing.
Adversarial examples contain carefully crafted perturbations that can fool deep neural networks (DNNs) into making wrong predictions. Enhancing the adversarial robustness of DNNs has gained considerable interest in recent years. Although image transformation-based defenses were widely considered at an earlier time, most of them have been defeated by adaptive attacks. In this paper, we propose a new image transformation defense based on error diffusion halftoning, and combine it with adversarial training to defend against adversarial examples. Error diffusion halftoning projects an image into a 1-bit space and diffuses quantization error to neighboring pixels. This process can remove adversarial perturbations from a given image while maintaining acceptable image quality in the meantime in favor of recognition. Experimental results demonstrate that the proposed method is able to improve adversarial robustness even under advanced adaptive attacks, while most of the other image transformation-based defenses do not. We show that a proper image transformation can still be an effective defense approach. Code: https://github.com/shaoyuanlo/Halftoning-Defense
The last few years have seen gigantic leaps in algorithms and systems to support efficient deep learning inference. Pruning and quantization algorithms can now consistently compress neural networks by an order of magnitude. For a compressed neural network, a multitude of inference frameworks have been designed to maximize the performance of the target hardware. While we find mature support for quantized neural networks in production frameworks such as OpenVINO and MNN, support for pruned sparse neural networks is still lacking. To tackle this challenge, we present SparseDNN, a sparse deep learning inference engine targeting CPUs. We present both kernel-level optimizations with a sparse code generator to accelerate sparse operators and novel network-level optimizations catering to sparse networks. We show that our sparse code generator can achieve significant speedups over state-of-the-art sparse and dense libraries. On end-to-end benchmarks such as Huggingface pruneBERT, SparseDNN achieves up to 5x throughput improvement over dense inference with state-of-the-art OpenVINO. Open source library at: https://github.com/marsupialtail/sparsednn.
In this paper, we approach the problem of optimizing blackbox functions over large hybrid search spaces consisting of both combinatorial and continuous parameters. We demonstrate that previous evolutionary algorithms which rely on mutation-based approaches, while flexible over combinatorial spaces, suffer from a curse of dimensionality in high dimensional continuous spaces both theoretically and empirically, which thus limits their scope over hybrid search spaces as well. In order to combat this curse, we propose ES-ENAS, a simple and modular joint optimization procedure combining the class of sample-efficient smoothed gradient techniques, commonly known as Evolutionary Strategies (ES), with combinatorial optimizers in a highly scalable and intuitive way, inspired by the one-shot or supernet paradigm introduced in Efficient Neural Architecture Search (ENAS). By doing so, we achieve significantly more sample efficiency, which we empirically demonstrate over synthetic benchmarks, and are further able to apply ES-ENAS for architecture search over popular RL benchmarks.
Deep learning models typically use single-precision (FP32) floating point data types for representing activations and weights, but a slew of recent research work has shown that computations with reduced-precision data types (FP16, 16-bit integers, 8-bit integers or even 4- or 2-bit integers) are enough to achieve same accuracy as FP32 and are much more efficient. Therefore, we designed fbgemm, a high-performance kernel library, from ground up to perform high-performance quantized inference on current generation CPUs. fbgemm achieves efficiency by fusing common quantization operations with a high-performance gemm implementation and by shape- and size-specific kernel code generation at runtime. The library has been deployed at Facebook, where it delivers greater than 2x performance gains with respect to our current production baseline.
We introduce an automated tool for deploying ultra low-latency, low-power deep neural networks with convolutional layers on FPGAs. By extending the hls4ml library, we demonstrate an inference latency of $5\,\mu$s using convolutional architectures, targeting microsecond latency applications like those at the CERN Large Hadron Collider. Considering benchmark models trained on the Street View House Numbers Dataset, we demonstrate various methods for model compression in order to fit the computational constraints of a typical FPGA device used in trigger and data acquisition systems of particle detectors. In particular, we discuss pruning and quantization-aware training, and demonstrate how resource utilization can be significantly reduced with little to no loss in model accuracy. We show that the FPGA critical resource consumption can be reduced by 97% with zero loss in model accuracy, and by 99% when tolerating a 6% accuracy degradation.
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.
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.
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.
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.
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.
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.
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.
This paper introduces neural architecture search (NAS) for the automatic discovery of small models for keyword spotting (KWS) in limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) to maximize the classification accuracy while minimizing the number of operations per inference. Using NAS only, we were able to obtain a highly efficient model with 95.4% accuracy on the Google speech commands dataset with 494.8 kB of memory usage and 19.6 million operations. Additionally, weight quantization is used to reduce the memory consumption even further. We show that weight quantization to low bit-widths (e.g. 1 bit) can be used without substantial loss in accuracy. By increasing the number of input features from 10 MFCC to 20 MFCC we were able to increase the accuracy to 96.3% at 340.1 kB of memory usage and 27.1 million operations.
Modern Neural Networks are eminent in achieving state of the art performance on tasks under Computer Vision, Natural Language Processing and related verticals. However, they are notorious for their voracious memory and compute appetite which further obstructs their deployment on resource limited edge devices. In order to achieve edge deployment, researchers have developed pruning and quantization algorithms to compress such networks without compromising their efficacy. Such compression algorithms are broadly experimented on standalone CNN and RNN architectures while in this work, we present an unconventional end to end compression pipeline of a CNN-LSTM based Image Captioning model. The model is trained using VGG16 or ResNet50 as an encoder and an LSTM decoder on the flickr8k dataset. We then examine the effects of different compression architectures on the model and design a compression architecture that achieves a 73.1% reduction in model size, 71.3% reduction in inference time and a 7.7% increase in BLEU score as compared to its uncompressed counterpart.
Neural network compression methods like pruning and quantization are very effective at efficiently deploying Deep Neural Networks (DNNs) on edge devices. However, DNNs remain vulnerable to adversarial examples-inconspicuous inputs that are specifically designed to fool these models. In particular, Universal Adversarial Perturbations (UAPs), are a powerful class of adversarial attacks which create adversarial perturbations that can generalize across a large set of inputs. In this work, we analyze the effect of various compression techniques to UAP attacks, including different forms of pruning and quantization. We test the robustness of compressed models to white-box and transfer attacks, comparing them with their uncompressed counterparts on CIFAR-10 and SVHN datasets. Our evaluations reveal clear differences between pruning methods, including Soft Filter and Post-training Pruning. We observe that UAP transfer attacks between pruned and full models are limited, suggesting that the systemic vulnerabilities across these models are different. This finding has practical implications as using different compression techniques can blunt the effectiveness of black-box transfer attacks. We show that, in some scenarios, quantization can produce gradient-masking, giving a false sense of security. Finally, our results suggest that conclusions about the robustness of compressed models to UAP attacks is application dependent, observing different phenomena in the two datasets used in our experiments.
Deep neural networks (DNNs) have been extremely successful in solving many challenging AI tasks in natural language processing, speech recognition, and computer vision nowadays. However, DNNs are typically computation intensive, memory demanding, and power hungry, which significantly limits their usage on platforms with constrained resources. Therefore, a variety of compression techniques (e.g. quantization, pruning, and knowledge distillation) have been proposed to reduce the size and power consumption of DNNs. Blockwise knowledge distillation is one of the compression techniques that can effectively reduce the size of a highly complex DNN. However, it is not widely adopted due to its long training time. In this paper, we propose a novel parallel blockwise distillation algorithm to accelerate the distillation process of sophisticated DNNs. Our algorithm leverages local information to conduct independent blockwise distillation, utilizes depthwise separable layers as the efficient replacement block architecture, and properly addresses limiting factors (e.g. dependency, synchronization, and load balancing) that affect parallelism. The experimental results running on an AMD server with four Geforce RTX 2080Ti GPUs show that our algorithm can achieve 3x speedup plus 19% energy savings on VGG distillation, and 3.5x speedup plus 29% energy savings on ResNet distillation, both with negligible accuracy loss. The speedup of ResNet distillation can be further improved to 3.87 when using four RTX6000 GPUs in a distributed cluster.
We propose an original method for vectorizing an image or zooming it at an arbitrary scale. The core of our method relies on the resolution of a geometric variational model and therefore offers theoretic guarantees. More precisely, it associates a total variation energy to every valid triangulation of the image pixels. Its minimization induces a trian-gulation that reflects image gradients. We then exploit this triangulation to precisely locate discontinuities, which can then simply be vectorized or zoomed. This new approach works on arbitrary images without any learning phase. It is particularly appealing for processing images with low quantization like pixel art and can be used for depixelizing such images. The method can be evaluated with an online demonstrator, where users can reproduce results presented here or upload their own images.
The paper presents a unified, flexible framework for the tasks of audio inpainting, declipping, and dequantization. The concept is further extended to cover analogous degradation models in a transformed domain, e.g. quantization of the signal's time-frequency coefficients. The task of reconstructing an audio signal from degraded observations in two different domains is formulated as an inverse problem, and several algorithmic solutions are developed. The viability of the presented concept is demonstrated on an example where audio reconstruction from partial and quantized observations of both the time-domain signal and its time-frequency coefficients is carried out.