Reproduce “Unifying Self-Supervised Clustering and Energy-Based Models” on OpenResearch

About this paper

Emanuele Sansone, Robin Manhaeve

GEDI presents a framework that unifies self-supervised clustering with energy-based models by providing a principled probabilistic interpretation of diverse SSL objectives. The approach demonstrates superior performance in clustering, data generation, and out-of-distribution detection, achieving these results while intrinsically avoiding common failure modes without relying on architectural heuristics.

Key insights

• Diverse self-supervised learning objectives can be understood and derived from a common probabilistic framework centered on maximizing log-likelihood lower bounds.
• The GEDI objective provides formal theoretical guarantees against the triad of common SSL failure modes, eliminating the need for ad-hoc architectural components or asymmetrical training procedures.
• A single network can be jointly trained for both discriminative representation learning and likelihood-based generative modeling, leveraging the strengths of both paradigms for improved performance across tasks.

Summary by alphaXiv.