What makes for uniformity for non-contrastive self-supervised learning?

Recent advances in self-supervised learning (SSL) have made remarkable progress, especially for contrastive methods that target pulling two augmented views of one image together and pushing the views of all other images away. In this setting, negative pairs play a key role in avoiding collapsed representation. Recent studies, such as those on bootstrap your own latent (BYOL) and SimSiam, have surprisingly achieved a comparable performance even without contrasting negative samples. However, a basic theoretical issue for SSL arises: how can different SSL methods avoid collapsed representation, and is there a common design principle? In this study, we look deep into current non-contrastive SSL methods and analyze the key factors that avoid collapses. To achieve this goal, we present a new indicator of uniformity metric and study the local dynamics of the indicator to diagnose collapses in different scenarios. Moreover, we present some principles for choosing a good predictor, such that we can explicitly control the optimization process. Our theoretical analysis result is validated on some widely used benchmarks spanning different-scale datasets. We also compare recent SSL methods and analyze their commonalities in avoiding collapses and some ideas for future algorithm designs.