Thu, April 30, 2026
11:30 am - 12:25 pm
Math Tower (MW) 154
Wonjun Lee
The Ohio State University
Title
Linear Separability in Contrastive Learning via Gradient Flow Dynamics
Abstract
Despite its widespread success in contrastive learning, the underlying mechanisms driving SimCLR are not fully understood. This talk presents a mathematical analysis that clarifies how the geometry of the learned latent distribution arises. We demonstrate that, despite a nonconvex loss function, gradient flow dynamics naturally guide training toward favorable representations by inducing early clustering in the feature space. Furthermore, we prove that under certain structural assumptions, these learned features become linearly separable with respect to ground truth labels. Numerical results supporting these theoretical insights will also be discussed.