{"title":"用于对比式自我监督学习的可瘦网络","authors":"Shuai Zhao, Linchao Zhu, Xiaohan Wang, Yi Yang","doi":"10.1007/s11263-024-02211-7","DOIUrl":null,"url":null,"abstract":"<p>Self-supervised learning makes significant progress in pre-training large models, but struggles with small models. Mainstream solutions to this problem rely mainly on knowledge distillation, which involves a two-stage procedure: first training a large teacher model and then distilling it to improve the generalization ability of smaller ones. In this work, we introduce another one-stage solution to obtain pre-trained small models without the need for extra teachers, namely, slimmable networks for contrastive self-supervised learning (SlimCLR). A slimmable network consists of a full network and several weight-sharing sub-networks, which can be pre-trained once to obtain various networks, including small ones with low computation costs. However, interference between weight-sharing networks leads to severe performance degradation in self-supervised cases, as evidenced by <i>gradient magnitude imbalance</i> and <i>gradient direction divergence</i>. The former indicates that a small proportion of parameters produce dominant gradients during backpropagation, while the main parameters may not be fully optimized. The latter shows that the gradient direction is disordered, and the optimization process is unstable. To address these issues, we introduce three techniques to make the main parameters produce dominant gradients and sub-networks have consistent outputs. These techniques include slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Furthermore, theoretical results are presented to demonstrate that a single slimmable linear layer is sub-optimal during linear evaluation. Thus a switchable linear probe layer is applied during linear evaluation. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs. The code is available at https://github.com/mzhaoshuai/SlimCLR.</p>","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"11 1","pages":""},"PeriodicalIF":11.6000,"publicationDate":"2024-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Slimmable Networks for Contrastive Self-supervised Learning\",\"authors\":\"Shuai Zhao, Linchao Zhu, Xiaohan Wang, Yi Yang\",\"doi\":\"10.1007/s11263-024-02211-7\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p>Self-supervised learning makes significant progress in pre-training large models, but struggles with small models. Mainstream solutions to this problem rely mainly on knowledge distillation, which involves a two-stage procedure: first training a large teacher model and then distilling it to improve the generalization ability of smaller ones. In this work, we introduce another one-stage solution to obtain pre-trained small models without the need for extra teachers, namely, slimmable networks for contrastive self-supervised learning (SlimCLR). A slimmable network consists of a full network and several weight-sharing sub-networks, which can be pre-trained once to obtain various networks, including small ones with low computation costs. However, interference between weight-sharing networks leads to severe performance degradation in self-supervised cases, as evidenced by <i>gradient magnitude imbalance</i> and <i>gradient direction divergence</i>. The former indicates that a small proportion of parameters produce dominant gradients during backpropagation, while the main parameters may not be fully optimized. The latter shows that the gradient direction is disordered, and the optimization process is unstable. To address these issues, we introduce three techniques to make the main parameters produce dominant gradients and sub-networks have consistent outputs. These techniques include slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Furthermore, theoretical results are presented to demonstrate that a single slimmable linear layer is sub-optimal during linear evaluation. Thus a switchable linear probe layer is applied during linear evaluation. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs. The code is available at https://github.com/mzhaoshuai/SlimCLR.</p>\",\"PeriodicalId\":13752,\"journal\":{\"name\":\"International Journal of Computer Vision\",\"volume\":\"11 1\",\"pages\":\"\"},\"PeriodicalIF\":11.6000,\"publicationDate\":\"2024-09-26\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"International Journal of Computer Vision\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://doi.org/10.1007/s11263-024-02211-7\",\"RegionNum\":2,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"International Journal of Computer Vision","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1007/s11263-024-02211-7","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
Slimmable Networks for Contrastive Self-supervised Learning
Self-supervised learning makes significant progress in pre-training large models, but struggles with small models. Mainstream solutions to this problem rely mainly on knowledge distillation, which involves a two-stage procedure: first training a large teacher model and then distilling it to improve the generalization ability of smaller ones. In this work, we introduce another one-stage solution to obtain pre-trained small models without the need for extra teachers, namely, slimmable networks for contrastive self-supervised learning (SlimCLR). A slimmable network consists of a full network and several weight-sharing sub-networks, which can be pre-trained once to obtain various networks, including small ones with low computation costs. However, interference between weight-sharing networks leads to severe performance degradation in self-supervised cases, as evidenced by gradient magnitude imbalance and gradient direction divergence. The former indicates that a small proportion of parameters produce dominant gradients during backpropagation, while the main parameters may not be fully optimized. The latter shows that the gradient direction is disordered, and the optimization process is unstable. To address these issues, we introduce three techniques to make the main parameters produce dominant gradients and sub-networks have consistent outputs. These techniques include slow start training of sub-networks, online distillation, and loss re-weighting according to model sizes. Furthermore, theoretical results are presented to demonstrate that a single slimmable linear layer is sub-optimal during linear evaluation. Thus a switchable linear probe layer is applied during linear evaluation. We instantiate SlimCLR with typical contrastive learning frameworks and achieve better performance than previous arts with fewer parameters and FLOPs. The code is available at https://github.com/mzhaoshuai/SlimCLR.
期刊介绍:
The International Journal of Computer Vision (IJCV) serves as a platform for sharing new research findings in the rapidly growing field of computer vision. It publishes 12 issues annually and presents high-quality, original contributions to the science and engineering of computer vision. The journal encompasses various types of articles to cater to different research outputs.
Regular articles, which span up to 25 journal pages, focus on significant technical advancements that are of broad interest to the field. These articles showcase substantial progress in computer vision.
Short articles, limited to 10 pages, offer a swift publication path for novel research outcomes. They provide a quicker means for sharing new findings with the computer vision community.
Survey articles, comprising up to 30 pages, offer critical evaluations of the current state of the art in computer vision or offer tutorial presentations of relevant topics. These articles provide comprehensive and insightful overviews of specific subject areas.
In addition to technical articles, the journal also includes book reviews, position papers, and editorials by prominent scientific figures. These contributions serve to complement the technical content and provide valuable perspectives.
The journal encourages authors to include supplementary material online, such as images, video sequences, data sets, and software. This additional material enhances the understanding and reproducibility of the published research.
Overall, the International Journal of Computer Vision is a comprehensive publication that caters to researchers in this rapidly growing field. It covers a range of article types, offers additional online resources, and facilitates the dissemination of impactful research.