With the proliferation of deep convolutional neural network (CNN) algorithms for mobile processing, limited precision quantization has become an essential tool for CNN efficiency. Consequently, various works have sought to design fixed precision quantization algorithms and quantization-focused optimization techniques that minimize quantization induced performance degradation. However, there is little concrete understanding of how various CNN design decisions/best practices affect quantized inference behaviour. Weight initialization strategies are often associated with solving issues such as vanishing/exploding gradients but an often-overlooked aspect is their impact on the final trained distributions of each layer. We present an in-depth, fine-grained ablation study of the effect of different weights initializations on the final distributions of weights and activations of different CNN architectures. The fine-grained, layerwise analysis enables us to gain deep insights on how initial weights distributions will affect final accuracy and quantized behaviour. To our best knowledge, we are the first to perform such a low-level, in-depth quantitative analysis of weights initialization and its effect on quantized behaviour.
Binary neural networks improve computationally efficiency of deep models with a large margin. However, there is still a performance gap between a successful full-precision training and binary training. We bring some insights about why this accuracy drop exists and call for a better understanding of binary network geometry. We start with analyzing full-precision neural networks with ReLU activation and compare it with its binarized version. This comparison suggests to initialize networks with random bias, a counter-intuitive remedy.
University of Maryland, Baltimore County（马里兰大学巴尔的摩县分校）
We present "Amenable Sparse Network Investigator" (ASNI) algorithm that utilizes a novel pruning strategy based on a sigmoid function that induces sparsity level globally over the course of one single round of training. The ASNI algorithm fulfills both tasks that current state-of-the-art strategies can only do one of them. The ASNI algorithm has two subalgorithms: 1) ASNI-I, 2) ASNI-II. ASNI-I learns an accurate sparse off-the-shelf network only in one single round of training. ASNI-II learns a sparse network and an initialization that is quantized, compressed, and from which the sparse networ…
An oft-cited challenge of federated learning is the presence of heterogeneity. \emph{Data heterogeneity} refers to the fact that data from different clients may follow very different distributions. \emph{System heterogeneity} refers to client devices having different system capabilities. A considerable number of federated optimization methods address this challenge. In the literature, empirical evaluations usually start federated training from random initialization. However, in many practical applications of federated learning, the server has access to proxy data for the training task that can…
Reducing the precision of weights and activation functions in neural network training, with minimal impact on performance, is essential for the deployment of these models in resource-constrained environments. We apply mean-field techniques to networks with quantized activations in order to evaluate the degree to which quantization degrades signal propagation at initialization. We derive initialization schemes which maximize signal propagation in such networks and suggest why this is helpful for generalization. Building on these results, we obtain a closed form implicit equation for $L_{\max}$,…
An oft-cited challenge of federated learning is the presence of heterogeneity. \emph{Data heterogeneity} refers to the fact that data from different clients may follow very different distributions. \emph{System heterogeneity} refers to the fact that client devices have different system capabilities. A considerable number of federated optimization methods address this challenge. In the literature, empirical evaluations usually start federated training from random initialization. However, in many practical applications of federated learning, the server has access to proxy data for the training t…
Normalization layers are a staple in state-of-the-art deep neural network architectures. They are widely believed to stabilize training, enable higher learning rate, accelerate convergence and improve generalization, though the reason for their effectiveness is still an active research topic. In this work, we challenge the commonly-held beliefs by showing that none of the perceived benefits is unique to normalization. Specifically, we propose fixed-update initialization (Fixup), an initialization motivated by solving the exploding and vanishing gradient problem at the beginning of training via…
Improving Learning of New Diseases through Knowledge-Enhanced Initialization for Federated Adapter Tuning
Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR)（高性能计算研究所（IHPC），科技研究局（A*STAR））；EVYD Technology（EVYD技术）
When large scale training data is available, one can obtain compact and accurate networks to be deployed in resource-constrained environments effectively through quantization and pruning. However, training data are often protected due to privacy concerns and it is challenging to obtain compact networks without data. We study data-free quantization and pruning by transferring knowledge from trained large networks to compact networks. Auxiliary generators are simultaneously and adversarially trained with the targeted compact networks to generate synthetic inputs that maximize the discrepancy bet…
Post-training dropout based approaches achieve high sparsity and are well established means of deciphering problems relating to computational cost and overfitting in Neural Network architectures. Contrastingly, pruning at initialization is still far behind. Initialization pruning is more efficacious when it comes to scaling computation cost of the network. Furthermore, it handles overfitting just as well as post training dropout. In approbation of the above reasons, the paper presents two approaches to prune at initialization. The goal is to achieve higher sparsity while preserving performance…
DoTA: Weight-Decomposed Tensor Adaptation for Large Language Models
Gaoling School of Artificial Intelligence, Renmin University of China（中国人民大学高瓴人工智能学院）；University of International Business and Economics, China（对外经济贸易大学）；XiaoMi AI Lab, China（小米AI实验室）
Seoul National University（首尔国立大学）；Qualcomm AI Research（高通人工智能研究院）；Qualcomm Technologies, Inc.（高通技术公司）
Quantization and Knowledge distillation (KD) methods are widely used to reduce memory and power consumption of deep neural networks (DNNs), especially for resource-constrained edge devices. Although their combination is quite promising to meet these requirements, it may not work as desired. It is mainly because the regularization effect of KD further diminishes the already reduced representation power of a quantized model. To address this short-coming, we propose Quantization-aware Knowledge Distillation (QKD) wherein quantization and KD are care-fully coordinated in three phases. First, Self-…
The ability to dynamically adapt neural networks to newly-available data without performance deterioration would revolutionize deep learning applications. Streaming learning (i.e., learning from one data example at a time) has the potential to enable such real-time adaptation, but current approaches i) freeze a majority of network parameters during streaming and ii) are dependent upon offline, base initialization procedures over large subsets of data, which damages performance and limits applicability. To mitigate these shortcomings, we propose Cold Start Streaming Learning (CSSL), a simple, e…
Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning
Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）；School of Artificial Intelligence, University of Chinese Academy of Sciences（中国科学院大学人工智能学院）；Zhongguancun Academy（中关村学院）；City University of Hong Kong（香港城市大学）；NJUST（南京理工大学）
Assigning Distinct Roles to Quantized and Low-Rank Matrices Toward Optimal Weight Decomposition
Department of Computer Science, Yonsei University（yonsei大学计算机科学系）；Department of Artificial Intelligence, Yonsei University（yonsei大学人工智能系）；Department of Electronic & Electrical Convergence Engineering, Hongik University（hongik大学电子与电气融合工程系）
A Fast Initialization Method for Neural Network Controllers: A Case Study of Image-based Visual Servoing Control for the multicopter Interception
State Key Laboratory of Fluid Power and Mechatronic Systems, School of Mechanical Engineering, Zhejiang University, Hangzhou, China（流体动力与机械系统国家重点实验室，机械工程学院，浙江大学，杭州，中国）；Zhiyuan Research Institute, Hangzhou, China（智远研究院，杭州，中国）
One of the difficulties of training deep neural networks is caused by improper scaling between layers. Scaling issues introduce exploding / gradient problems, and have typically been addressed by careful scale-preserving initialization. We investigate the value of preserving scale, or isometry, beyond the initial weights. We propose two methods of maintaing isometry, one exact and one stochastic. Preliminary experiments show that for both determinant and scale-normalization effectively speeds up learning. Results suggest that isometry is important in the beginning of learning, and maintaining …
OnlineQAT: On-Policy Distillation for Ultra-Low-Bit Large Language Models
The Hong Kong Polytechnic University（香港理工大学）；Sun Yat-sen University（中山大学）；PolyU-Daya Bay Technology and Innovation Research Institute（香港理工大学大亚湾技术创新研究院）
School of Optoelectronic, Beijing Institute of Technology（北京理工大学光电学院）；Graduate School at Shenzhen, Tsinghua University（清华大学深圳研究生院）；BGI Research（华大研究院）
In this paper, we propose the nonlinearity generation method to speed up and stabilize the training of deep convolutional neural networks. The proposed method modifies a family of activation functions as nonlinearity generators (NGs). NGs make the activation functions linear symmetric for their inputs to lower model capacity, and automatically introduce nonlinearity to enhance the capacity of the model during training. The proposed method can be considered an unusual form of regularization: the model parameters are obtained by training a relatively low-capacity model, that is relatively easy t…
HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs
School of Software and Microelectronics, Peking University, Beijing, China（软件与微电子学院，北京大学，北京，中国）；School of Computer Science, Peking University, Beijing, China（计算机科学学院，北京大学，北京，中国）
Training deep neural networks requires intricate initialization and careful selection of learning rates. The emergence of stochastic gradient optimization methods that use adaptive learning rates based on squared past gradients, e.g., AdaGrad, AdaDelta, and Adam, eases the job slightly. However, such methods have also been proven problematic in recent studies with their own pitfalls including non-convergence issues and so on. Alternative variants have been proposed for enhancement, such as AMSGrad, AdaShift and AdaBound. In this work, we identify a new problem of adaptive learning rate methods…
