Curriculum-Learned Masked Autoencoders (CL-MAE): Enhancing Autoencoders through Curriculum Learning

[Submitted on 31 Aug 2023]

Download a PDF of the paper titled CL-MAE: Curriculum-Learned Masked Autoencoders, by Neelu Madan and 4 other authors

Download PDF

Abstract: Masked image modeling has been proven effective in generating robust representations for multiple downstream tasks. This paper introduces a curriculum learning approach called CL-MAE that updates the masking strategy to increase the complexity of the self-supervised reconstruction task. A learnable masking module is integrated into masked autoencoders (MAE), which gradually transitions from a partner to an adversary during training. The resulting CL-MAE exhibits superior representation learning capabilities compared to MAE, as confirmed by empirical results on five downstream tasks. Curriculum learning can be successfully used to self-supervise masked autoencoders.