Vision Transformers (ViTs) have achieved state-of-the-art performance on various vision tasks. However, ViTs' self-attention module is still arguably a major bottleneck, limiting their achievable hardware efficiency. Meanwhile, existing accelerators dedicated to NLP Transformers are not optimal for ViTs. This is because there is a large difference between ViTs and NLP Transformers: ViTs have a relatively fixed number of input tokens, whose attention maps can be pruned by up to 90% even with fixed sparse patterns; while NLP Transformers need to handle input sequences of varying numbers of tokens and rely on on-the-fly predictions of dynamic sparse attention patterns for each input to achieve a decent sparsity (e.g., >=50%). To this end, we propose a dedicated algorithm and accelerator co-design framework dubbed ViTCoD for accelerating ViTs. Specifically, on the algorithm level, ViTCoD prunes and polarizes the attention maps to have either denser or sparser fixed patterns for regularizing two levels of workloads without hurting the accuracy, largely reducing the attention computations while leaving room for alleviating the remaining dominant data movements; on top of that, we further integrate a lightweight and learnable auto-encoder module to enable trading the dominant high-cost data movements for lower-cost computations. On the hardware level, we develop a dedicated accelerator to simultaneously coordinate the enforced denser/sparser workloads and encoder/decoder engines for boosted hardware utilization. Extensive experiments and ablation studies validate that ViTCoD largely reduces the dominant data movement costs, achieving speedups of up to 235.3x, 142.9x, 86.0x, 10.1x, and 6.8x over general computing platforms CPUs, EdgeGPUs, GPUs, and prior-art Transformer accelerators SpAtten and Sanger under an attention sparsity of 90%, respectively.
相關內容
Fine-tuning pre-trained models has been ubiquitously proven to be effective in a wide range of NLP tasks. However, fine-tuning the whole model is parameter inefficient as it always yields an entirely new model for each task. Currently, many research works propose to only fine-tune a small portion of the parameters while keeping most of the parameters shared across different tasks. These methods achieve surprisingly good performance and are shown to be more stable than their corresponding fully fine-tuned counterparts. However, such kind of methods is still not well understood. Some natural questions arise: How does the parameter sparsity lead to promising performance? Why is the model more stable than the fully fine-tuned models? How to choose the tunable parameters? In this paper, we first categorize the existing methods into random approaches, rule-based approaches, and projection-based approaches based on how they choose which parameters to tune. Then, we show that all of the methods are actually sparse fine-tuned models and conduct a novel theoretical analysis of them. We indicate that the sparsity is actually imposing a regularization on the original model by controlling the upper bound of the stability. Such stability leads to better generalization capability which has been empirically observed in a lot of recent research works. Despite the effectiveness of sparsity grounded by our theory, it still remains an open problem of how to choose the tunable parameters. To better choose the tunable parameters, we propose a novel Second-order Approximation Method (SAM) which approximates the original problem with an analytically solvable optimization function. The tunable parameters are determined by directly optimizing the approximation function. The experimental results show that our proposed SAM model outperforms many strong baseline models and it also verifies our theoretical analysis.
Low-rankness plays an important role in traditional machine learning, but is not so popular in deep learning. Most previous low-rank network compression methods compress the networks by approximating pre-trained models and re-training. However, the optimal solution in the Euclidean space may be quite different from the one in the low-rank manifold. A well-pre-trained model is not a good initialization for the model with low-rank constraints. Thus, the performance of a low-rank compressed network degrades significantly. Compared to other network compression methods such as pruning, low-rank methods attracts less attention in recent years. In this paper, we devise a new training method, low-rank projection with energy transfer (LRPET), that trains low-rank compressed networks from scratch and achieves competitive performance. First, we propose to alternately perform stochastic gradient descent training and projection onto the low-rank manifold. Compared to re-training on the compact model, this enables full utilization of model capacity since solution space is relaxed back to Euclidean space after projection. Second, the matrix energy (the sum of squares of singular values) reduction caused by projection is compensated by energy transfer. We uniformly transfer the energy of the pruned singular values to the remaining ones. We theoretically show that energy transfer eases the trend of gradient vanishing caused by projection. Third, we propose batch normalization (BN) rectification to cut off its effect on the optimal low-rank approximation of the weight matrix, which further improves the performance. Comprehensive experiments on CIFAR-10 and ImageNet have justified that our method is superior to other low-rank compression methods and also outperforms recent state-of-the-art pruning methods. Our code is available at //github.com/BZQLin/LRPET.
Deep reinforcement learning has achieved great success in various fields with its super decision-making ability. However, the policy learning process requires a large amount of training time, causing energy consumption. Inspired by the redundancy of neural networks, we propose a lightweight parallel training framework based on neural network compression, AcceRL, to accelerate the policy learning while ensuring policy quality. Specifically, AcceRL speeds up the experience collection by flexibly combining various neural network compression methods. Overall, the AcceRL consists of five components, namely Actor, Learner, Compressor, Corrector, and Monitor. The Actor uses the Compressor to compress the Learner's policy network to interact with the environment. And the generated experiences are transformed by the Corrector with Off-Policy methods, such as V-trace, Retrace and so on. Then the corrected experiences are feed to the Learner for policy learning. We believe this is the first general reinforcement learning framework that incorporates multiple neural network compression techniques. Extensive experiments conducted in gym show that the AcceRL reduces the time cost of the actor by about 2.0 X to 4.13 X compared to the traditional methods. Furthermore, the AcceRL reduces the whole training time by about 29.8% to 40.3% compared to the traditional methods while keeps the same policy quality.
To efficiently exploit the massive amounts of raw data that are increasingly being generated in mobile edge networks, federated learning (FL) has emerged as a promising distributed learning technique. By collaboratively training a shared learning model on edge devices, raw data transmission and storage are replaced by the exchange of the local computed parameters/gradients in FL, which thus helps address latency and privacy issues. However, the number of resource blocks when using traditional orthogonal transmission strategies for FL linearly scales with the number of participating devices, which conflicts with the scarcity of communication resources. To tackle this issue, over-the-air computation (AirComp) has emerged recently which leverages the inherent superposition property of wireless channels to perform one-shot model aggregation. However, the aggregation accuracy in AirComp suffers from the unfavorable wireless propagation environment. In this paper, we consider the use of intelligent reflecting surfaces (IRSs) to mitigate this problem and improve FL performance with AirComp. Specifically, a performance-oriented design scheme that directly minimizes the optimality gap of the loss function is proposed to accelerate the convergence of AirComp-based FL. We first analyze the convergence behavior of the FL procedure with the absence of channel fading and noise. Based on the obtained optimality gap which characterizes the impact of channel fading and noise in different communication rounds on the ultimate performance of FL, we propose both online and offline approaches to tackle the resulting design problem. Simulation results demonstrate that such a performance-oriented design strategy can achieve higher test accuracy than the conventional isolated mean square error (MSE) minimization approach in FL.
In recent years, group buying has become one popular kind of online shopping activity, thanks to its larger sales and lower unit price. Unfortunately, research seldom focuses on recommendations specifically for group buying by now. Although some recommendation models have been proposed for group recommendation, they can not be directly used to achieve real-world group buying recommendation, due to the essential difference between group recommendation and group buying recommendation. In this paper, we first formalize the task of group buying recommendations into two sub-tasks. Then, based on our insights into the correlations and interactions between the two sub-tasks, we propose a novel recommendation model for group buying, MGBR, built mainly with a multi-task learning module. To improve recommendation performance further, we devise some collaborative expert networks and adjusted gates in the multi-task learning module, to promote the information interaction between the two sub-tasks. Furthermore, we propose two auxiliary losses corresponding to the two sub-tasks, to refine the representation learning in our model. Our extensive experiments not only demonstrate that the augmented representations in our model result in better performance than previous recommendation models, but also justify the impacts of the specially designed components in our model.
Most deep learning methods for video frame interpolation consist of three main components: feature extraction, motion estimation, and image synthesis. Existing approaches are mainly distinguishable in terms of how these modules are designed. However, when interpolating high-resolution images, e.g. at 4K, the design choices for achieving high accuracy within reasonable memory requirements are limited. The feature extraction layers help to compress the input and extract relevant information for the latter stages, such as motion estimation. However, these layers are often costly in parameters, computation time, and memory. We show how ideas from dimensionality reduction combined with a lightweight optimization can be used to compress the input representation while keeping the extracted information suitable for frame interpolation. Further, we require neither a pretrained flow network nor a synthesis network, additionally reducing the number of trainable parameters and required memory. When evaluating on three 4K benchmarks, we achieve state-of-the-art image quality among the methods without pretrained flow while having the lowest network complexity and memory requirements overall.
Fully homomorphic encryption (FHE) protects data privacy in cloud computing by enabling computations to directly occur on ciphertexts. Although the speed of computationally expensive FHE operations can be significantly boosted by prior ASIC-based FHE accelerators, the performance of key-switching, the dominate primitive in various FHE operations, is seriously limited by their small bit-width datapaths and frequent matrix transpositions. In this paper, we present an electro-optical (EO) FHE accelerator, CryptoLight, to accelerate FHE operations. Its 512-bit datapath supporting 510-bit residues greatly reduces the key-switching cost. We also create an in-scratchpad-memory transpose unit to fast transpose matrices. Compared to prior FHE accelerators, on average, CryptoLight reduces the latency of various FHE applications by >94.4% and the energy consumption by >95%.
For time-critical IoT applications using deep learning, inference acceleration through distributed computing is a promising approach to meet a stringent deadline. In this paper, we implement a working prototype of a new distributed inference acceleration method HALP using three raspberry Pi 4. HALP accelerates inference by designing a seamless collaboration among edge devices (EDs) in Edge Computing. We maximize the parallelization between communication and computation among the collaborative EDs by optimizing the task partitioning ratio based on the segment-based partitioning. Experimental results show that the distributed inference HALP achieves 1.7x inference acceleration for VGG-16. Then, we combine distributed inference with conventional neural network model compression by setting up different shrinking hyperparameters for MobileNet-V1. In this way, we can further accelerate inference but at the cost of inference accuracy loss. To strike a balance between latency and accuracy, we propose dynamic model selection to select a model which provides the highest accuracy within the latency constraint. It is shown that the model selection with distributed inference HALP can significantly improve service reliability compared to the conventional stand-alone computation.
Privacy and security have rapidly emerged as priorities in system design. One powerful solution for providing both is privacy-preserving computation, where functions are computed directly on encrypted data and control can be provided over how data is used. Garbled circuits (GCs) are a PPC technology that provide both confidential computing and control over how data is used. The challenge is that they incur significant performance overheads compared to plaintext. This paper proposes a novel garbled circuit accelerator and compiler, named HAAC, to mitigate performance overheads and make privacy-preserving computation more practical. HAAC is a hardware-software co-design. GCs are exemplars of co-design as programs are completely known at compile time, i.e., all dependence, memory accesses, and control flow are fixed. The design philosophy of HAAC is to keep hardware simple and efficient, maximizing area devoted to our proposed custom execution units and other circuits essential for high performance (e.g., on-chip storage). The compiler can leverage its program understanding to realize hardware's performance potential by generating effective instruction schedules, data layouts, and orchestrating off-chip events. In taking this approach we can achieve ASIC performance/efficiency without sacrificing generality. Insights of our approach include how co-design enables expressing arbitrary GC programs as streams, which simplifies hardware and enables complete memory-compute decoupling, and the development of a scratchpad that captures data reuse by tracking program execution, eliminating the need for costly hardware managed caches and tagging logic. We evaluate HAAC with VIP-Bench and achieve a speedup of 608$\times$ in 4.3mm$^2$ of area.
Transformers have achieved superior performances in many tasks in natural language processing and computer vision, which also intrigues great interests in the time series community. Among multiple advantages of transformers, the ability to capture long-range dependencies and interactions is especially attractive for time series modeling, leading to exciting progress in various time series applications. In this paper, we systematically review transformer schemes for time series modeling by highlighting their strengths as well as limitations through a new taxonomy to summarize existing time series transformers in two perspectives. From the perspective of network modifications, we summarize the adaptations of module level and architecture level of the time series transformers. From the perspective of applications, we categorize time series transformers based on common tasks including forecasting, anomaly detection, and classification. Empirically, we perform robust analysis, model size analysis, and seasonal-trend decomposition analysis to study how Transformers perform in time series. Finally, we discuss and suggest future directions to provide useful research guidance. To the best of our knowledge, this paper is the first work to comprehensively and systematically summarize the recent advances of Transformers for modeling time series data. We hope this survey will ignite further research interests in time series Transformers.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191