Understanding self-supervised Learning Dynamics without Contrastive Pairs

TL;DR

This paper provides a theoretical understanding of how non-contrastive self-supervised learning methods avoid collapse and introduces DirectPred, a simple predictor setting approach that matches or exceeds complex predictors in performance.

Contribution

The paper offers a theoretical analysis of non-contrastive SSL dynamics and proposes DirectPred, a predictor setting method that eliminates the need for gradient training of the predictor.

Findings

DirectPred performs comparably to complex predictors on ImageNet.

Theoretical insights explain how non-contrastive SSL avoids collapse.

The study aligns with ablation results on STL-10 and ImageNet.

Abstract

While contrastive approaches of self-supervised learning (SSL) learn representations by minimizing the distance between two augmented views of the same data point (positive pairs) and maximizing views from different data points (negative pairs), recent \emph{non-contrastive} SSL (e.g., BYOL and SimSiam) show remarkable performance {\it without} negative pairs, with an extra learnable predictor and a stop-gradient operation. A fundamental question arises: why do these methods not collapse into trivial representations? We answer this question via a simple theoretical study and propose a novel approach, DirectPred, that \emph{directly} sets the linear predictor based on the statistics of its inputs, without gradient training. On ImageNet, it performs comparably with more complex two-layer non-linear predictors that employ BatchNorm and outperforms a linear predictor by $2.5\%$ in 300-epoch training (and $5\%$ in 60-epoch). DirectPred is motivated by our theoretical study of the nonlinear learning dynamics of non-contrastive SSL in simple linear networks. Our study yields conceptual insights into how non-contrastive SSL methods learn, how they avoid representational collapse, and how multiple factors, like predictor networks, stop-gradients, exponential moving averages, and weight decay all come into play. Our simple theory recapitulates the results of real-world ablation studies in both STL-10 and ImageNet. Code is released https://github.com/facebookresearch/luckmatters/tree/master/ssl.