Doctoral Thesis: Contrastive Learning Through Ranking Models

By: Melihcan Erol

Details

• Date: Friday, April 17
• Time: 1:00 pm - 2:00 pm
• Location: Haus Room (36-428)

Additional Location Details:

Abstract:
InfoNCE is the standard loss in contrastive learning, but its softmax form is not just a computational convenience: it hides a statistical assumption about how the top-scoring example is chosen. We show that this assumption is not well aligned with the normalized embedding setting used in many contrastive learning applications. By interpreting InfoNCE as the negative log-likelihood of a Plackett–Luce choice model with Gumbel noise, we make explicit the tail geometry encoded by the softmax link and use extreme value theory to derive geometry-consistent alternatives. Based on this observation, we propose WEINCE, a simple modification of InfoNCE that adapts the loss to the bounded-score regime using online batch statistics and no additional trainable parameters. Across five vision benchmarks, including Tiny-ImageNet, and on SimCSE sentence embeddings, WEINCE yields consistent improvements in frozen feature evaluation. We further show that the same extreme value argument extends beyond contrastive learning to RLHF reward model training, where bounded sigmoid scores and top-1 candidate selection create an analogous mismatch. On the Nectar and UltraFeedback datasets, the corrected link improves ranking metrics and calibration, with gains that scale with the number of candidates as predicted by the theory. Our results suggest that accounting for extreme value tail geometry is a promising and underexplored direction for improving softmax-based selection objectives more broadly.

Host

• Melihcan Erol
• Email: hsmerol@mit.edu