Timezone: »
Masked Autoencoders (MAE) based on a reconstruction task have risen to be a promising paradigm for self-supervised learning (SSL) and achieve state-of-the-art performance across different benchmark datasets. However, despite its impressive empirical success, there is still limited theoretical understanding of it. In this paper, we propose a theoretical understanding of how masking matters for MAE to learn meaningful features. We establish a close connection between MAE and contrastive learning, which shows that MAE implicit aligns the mask-induced positive pairs. Built upon this connection, we develop the first downstream guarantees for MAE methods, and analyze the effect of mask ratio. Besides, as a result of the implicit alignment, we also point out the dimensional collapse issue of MAE, and propose a Uniformity-enhanced MAE (U-MAE) loss that can effectively address this issue and bring significant improvements on real-world datasets, including CIFAR-10, ImageNet-100, and ImageNet-1K. Code is available at https://github.com/zhangq327/U-MAE.
Author Information
Qi Zhang (Peking University)
Yifei Wang (Peking University)
Yisen Wang (Peking University)
More from the Same Authors
-
2022 Poster: Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors »
Qixun Wang · Yifei Wang · Hong Zhu · Yisen Wang -
2022 Poster: When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture »
Yichuan Mo · Dongxian Wu · Yifei Wang · Yiwen Guo · Yisen Wang -
2023 Poster: Balance, Imbalance, and Rebalance: Understanding Robust Overfitting from a Minimax Game Perspective »
Yifei Wang · Liangchen Li · Jiansheng Yang · Zhouchen Lin · Yisen Wang -
2023 Poster: Adversarial Examples Are Not Real Features »
Ang Li · Yifei Wang · Yiwen Guo · Yisen Wang -
2023 Poster: GEQ: Gaussian Kernel Inspired Equilibrium Models »
Mingjie Li · Yisen Wang · Zhouchen Lin -
2023 Poster: Architecture Matters: Uncovering Implicit Mechanisms in Graph Contrastive Learning »
Xiaojun Guo · Yifei Wang · Zeming Wei · Yisen Wang -
2023 Poster: Tri-contrastive Learning: Identifiable Representation Learning with Automatic Discovery of Feature Importance »
Qi Zhang · Yifei Wang · Yisen Wang -
2023 Poster: Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding »
George Ma · Yifei Wang · Yisen Wang -
2022 Spotlight: Lightning Talks 6A-2 »
Yichuan Mo · Botao Yu · Gang Li · Zezhong Xu · Haoran Wei · Arsene Fansi Tchango · Raef Bassily · Haoyu Lu · Qi Zhang · Songming Liu · Mingyu Ding · Peiling Lu · Yifei Wang · Xiang Li · Dongxian Wu · Ping Guo · Wen Zhang · Hao Zhongkai · Mehryar Mohri · Rishab Goel · Yisen Wang · Yifei Wang · Yangguang Zhu · Zhi Wen · Ananda Theertha Suresh · Chengyang Ying · Yujie Wang · Peng Ye · Rui Wang · Nanyi Fei · Hui Chen · Yiwen Guo · Wei Hu · Chenglong Liu · Julien Martel · Yuqi Huo · Wu Yichao · Hang Su · Yisen Wang · Peng Wang · Huajun Chen · Xu Tan · Jun Zhu · Ding Liang · Zhiwu Lu · Joumana Ghosn · Shanshan Zhang · Wei Ye · Ze Cheng · Shikun Zhang · Tao Qin · Tie-Yan Liu -
2022 Spotlight: How Mask Matters: Towards Theoretical Understandings of Masked Autoencoders »
Qi Zhang · Yifei Wang · Yisen Wang -
2022 Spotlight: When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecture »
Yichuan Mo · Dongxian Wu · Yifei Wang · Yiwen Guo · Yisen Wang -
2022 Spotlight: Lightning Talks 1B-3 »
Chaofei Wang · Qixun Wang · Jing Xu · Long-Kai Huang · Xi Weng · Fei Ye · Harsh Rangwani · shrinivas ramasubramanian · Yifei Wang · Qisen Yang · Xu Luo · Lei Huang · Adrian G. Bors · Ying Wei · Xinglin Pan · Sho Takemori · Hong Zhu · Rui Huang · Lei Zhao · Yisen Wang · Kato Takashi · Shiji Song · Yanan Li · Rao Anwer · Yuhei Umeda · Salman Khan · Gao Huang · Wenjie Pei · Fahad Shahbaz Khan · Venkatesh Babu R · Zenglin Xu -
2022 Spotlight: Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors »
Qixun Wang · Yifei Wang · Hong Zhu · Yisen Wang -
2021 Poster: Dissecting the Diffusion Process in Linear Graph Convolutional Networks »
Yifei Wang · Yisen Wang · Jiansheng Yang · Zhouchen Lin -
2021 Poster: Residual Relaxation for Multi-view Representation Learning »
Yifei Wang · Zhengyang Geng · Feng Jiang · Chuming Li · Yisen Wang · Jiansheng Yang · Zhouchen Lin -
2020 Poster: Adversarial Weight Perturbation Helps Robust Generalization »
Dongxian Wu · Shu-Tao Xia · Yisen Wang