🤖 AI Summary
This work addresses the long-overlooked role of embedding norm in self-supervised learning (SSL). Motivated by the paradox that most SSL methods enforce unit-norm embeddings while discarding norm information, we theoretically establish—for the first time—that embedding norm directly governs SSL convergence rate and encodes sample anomaly (e.g., small norms indicate out-of-distribution or corrupted samples). Methodologically, we propose a norm-sensitivity diagnostic tool and a normalization-decoupled training framework, validated through theoretical analysis, controlled simulations, and large-scale experiments on SimCLR and BYOL. Our key contributions are: (i) uncovering a dual regulatory mechanism wherein norm jointly controls convergence speed and model confidence; (ii) achieving up to 2.3× training acceleration; and (iii) improving anomaly detection AUC by 12.7%. This work challenges the cosine-similarity-dominated paradigm in SSL and establishes a new theoretical foundation and practical benchmark for interpretability and robustness.
📝 Abstract
Self-supervised learning (SSL) allows training data representations without a supervised signal and has become an important paradigm in machine learning. Most SSL methods employ the cosine similarity between embedding vectors and hence effectively embed data on a hypersphere. While this seemingly implies that embedding norms cannot play any role in SSL, a few recent works have suggested that embedding norms have properties related to network convergence and confidence. In this paper, we resolve this apparent contradiction and systematically establish the embedding norm's role in SSL training. Using theoretical analysis, simulations, and experiments, we show that embedding norms (i) govern SSL convergence rates and (ii) encode network confidence, with smaller norms corresponding to unexpected samples. Additionally, we show that manipulating embedding norms can have large effects on convergence speed. Our findings demonstrate that SSL embedding norms are integral to understanding and optimizing network behavior.