Supervised Contrastive Representation Learning: Landscape Analysis with Unconstrained Features

The author explores the distinctive Neural-collapse (NC) phenomenon in over-parameterized deep neural networks, focusing on the supervised contrastive (SC) loss and its implications. Through analytical methods, the study delves into the solutions derived from optimizing the SC loss.