Paper ID: 2405.01053

Explicitly Modeling Universality into Self-Supervised Learning

Jingyao Wang, Wenwen Qiang, Zeen Song, Lingyu Si, Jiangmeng Li, Changwen Zheng, Bing Su

The goal of universality in self-supervised learning (SSL) is to learn universal representations from unlabeled data and achieve excellent performance on all samples and tasks. However, these methods lack explicit modeling of the universality in the learning objective, and the related theoretical understanding remains limited. This may cause models to overfit in data-scarce situations and generalize poorly in real life. To address these issues, we provide a theoretical definition of universality in SSL, which constrains both the learning and evaluation universality of the SSL models from the perspective of discriminability, transferability, and generalization. Then, we propose a $\sigma$-measurement to help quantify the score of one SSL model's universality. Based on the definition and measurement, we propose a general SSL framework, called GeSSL, to explicitly model universality into SSL. It introduces a self-motivated target based on $\sigma$-measurement, which enables the model to find the optimal update direction towards universality. Extensive theoretical and empirical evaluations demonstrate the superior performance of GeSSL.

Submitted: May 2, 2024