Reproduce “Probabilistic Variational Contrastive Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minoh Jeong, Seonho Kim, Alfred Hero
Variational Contrastive Learning (VCL) introduces a probabilistic framework for contrastive representation learning, enabling uncertainty quantification for embeddings through a decoder-free Evidence Lower Bound (ELBO) maximization. The method consistently matches or exceeds the classification performance of deterministic baselines while effectively mitigating dimensional collapse and providing interpretable posterior uncertainty measures.
Key insights
- InfoNCE loss can be interpreted as a surrogate for the reconstruction term within a variational ELBO framework, providing a strong probabilistic foundation for contrastive learning without requiring a generative decoder.
- Posterior dispersion (quantified by log det(K)) serves as an interpretable measure, inversely correlating with typicality within a class and with label ambiguity or corruption severity.
- Explicitly enforcing a uniform prior on the hypersphere via KL regularization effectively mitigates dimensional collapse, promoting a more isotropic use of the embedding space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.