On Feature Decorrelation in Self-Supervised Learning

被引:39
|
作者
Hua, Tianyu [1 ,2 ]
Wang, Wenxiao [1 ]
Xue, Zihui [2 ,3 ]
Ren, Sucheng [2 ,5 ]
Wang, Yue [4 ]
Zhao, Hang [1 ,2 ]
机构
[1] Tsinghua Univ, Beijing, Peoples R China
[2] Shanghai Qi Zhi Inst, Shanghai, Peoples R China
[3] UT Austin, Austin, TX USA
[4] MIT, Cambridge, MA 02139 USA
[5] South China Univ Technol, Guangzhou, Peoples R China
关键词
D O I
10.1109/ICCV48922.2021.00946
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
In self-supervised representation learning, a common idea behind most of the state-of-the-art approaches is to enforce the robustness of the representations to predefined augmentations. A potential issue of this idea is the existence of completely collapsed solutions (i.e., constant features), which are typically avoided implicitly by carefully chosen implementation details. In this work, we study a relatively concise framework containing the most common components from recent approaches. We verify the existence of complete collapse and discover another reachable collapse pattern that is usually overlooked, namely dimensional collapse. We connect dimensional collapse with strong correlations between axes and consider such connection as a strong motivation for feature decorrelation (i.e., standardizing the covariance matrix). The gains from feature decorrelation are verified empirically to highlight the importance and the potential of this insight.
引用
收藏
页码:9578 / 9588
页数:11
相关论文
共 50 条
  • [1] Self-Supervised Feature Learning by Learning to Spot Artifacts
    Jenni, Simon
    Favaro, Paolo
    [J]. 2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2018, : 2733 - 2742
  • [2] Digging Into Self-Supervised Learning of Feature Descriptors
    Melekhov, Iaroslav
    Laskar, Zakaria
    Li, Xiaotian
    Wang, Shuzhe
    Kannala, Juho
    [J]. 2021 INTERNATIONAL CONFERENCE ON 3D VISION (3DV 2021), 2021, : 1144 - 1155
  • [3] FLSL: Feature-level Self-supervised Learning
    Su, Qing
    Netchaev, Anton
    Li, Hai
    Ji, Shihao
    [J]. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 36 (NEURIPS 2023), 2023,
  • [4] Concurrent Discrimination and Alignment for Self-Supervised Feature Learning
    Dutta, Anjan
    Mancini, Massimiliano
    Akata, Zeynep
    [J]. 2021 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS (ICCVW 2021), 2021, : 2189 - 2198
  • [5] Self-Supervised Representation Learning by Rotation Feature Decoupling
    Feng, Zeyu
    Xu, Chang
    Tao, Dacheng
    [J]. 2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2019), 2019, : 10356 - 10366
  • [6] Self-Supervised Learning Disentangled Group Representation as Feature
    Wang, Tan
    Yue, Zhongqi
    Huang, Jianqiang
    Sun, Qianru
    Zhang, Hanwang
    [J]. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 34 (NEURIPS 2021), 2021, 34
  • [7] FundusNet, A self-supervised contrastive learning framework for Fundus Feature Learning
    Mojab, Nooshin
    Alam, Minhaj
    Hallak, Joelle
    [J]. INVESTIGATIVE OPHTHALMOLOGY & VISUAL SCIENCE, 2022, 63 (07)
  • [8] Toward Understanding the Feature Learning Process of Self-supervised Contrastive Learning
    Wen, Zixin
    Li, Yuanzhi
    [J]. INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 139, 2021, 139
  • [9] Generalized Semi-Supervised Learning via Self-Supervised Feature Adaptation
    Liang, Jiachen
    Hou, Ruibing
    Chang, Hong
    Ma, Bingpeng
    Shan, Shiguang
    Chen, Xilin
    [J]. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 36 (NEURIPS 2023), 2023,
  • [10] Self-Supervised Feature Enhancement: Applying Internal Pretext Task to Supervised Learning
    Xie, Tianshu
    Yang, Yuhang
    Ding, Zilin
    Cheng, Xuan
    Wang, Xiaomin
    Gong, Haigang
    Liu, Ming
    [J]. IEEE ACCESS, 2023, 11 : 1708 - 1717