Improving Self-Supervised Learning by Characterizing Idealized Representations
This work addresses a foundational gap in understanding self-supervised learning for researchers, providing a framework to enhance existing methods and derive new ones, though it is incremental in building on prior SSL approaches.
The paper tackled the problem of unclear characteristics leading to high downstream accuracies in self-supervised learning by characterizing idealized properties for representations, resulting in improved SSL algorithms that outperform baselines on benchmarks like SwAV+multicrops on ImageNet linear probing.
Despite the empirical successes of self-supervised learning (SSL) methods, it is unclear what characteristics of their representations lead to high downstream accuracies. In this work, we characterize properties that SSL representations should ideally satisfy. Specifically, we prove necessary and sufficient conditions such that for any task invariant to given data augmentations, desired probes (e.g., linear or MLP) trained on that representation attain perfect accuracy. These requirements lead to a unifying conceptual framework for improving existing SSL methods and deriving new ones. For contrastive learning, our framework prescribes simple but significant improvements to previous methods such as using asymmetric projection heads. For non-contrastive learning, we use our framework to derive a simple and novel objective. Our resulting SSL algorithms outperform baselines on standard benchmarks, including SwAV+multicrops on linear probing of ImageNet.