Improve Self-supervision Learning by Enhancing Invariant Information

Xuhao Pan, Chunshi Wang, Bin Zhao, Huaji Wang

Abstract: Self-supervised learning framework based on redundancy reduction has achieved remarkable success in the visual field and demonstrated excellent performance in various downstream tasks. However, current methods are largely constrained by a singular image-level perturbation range, which we believe limits the model’s ability to learn invariant information at different levels. Therefore, we explore a broader perturbation space and propose a new self-supervised learning method named enhanced-invariance Barlow Twins (EIBT). This method introduces additional feature-level perturbation strategy to provide more invariant information for the network’s learning, effectively enhancing the representational capacity of the network. Experimental results demonstrate that EIBT achieves the state-of-the-art performance across multiple benchmark datasets. The codes of our work can be found in https://github.com/2L84/EIBT-main.

External IDs:dblp:conf/icic/PanWZW25