Search Papers | Poster Sessions | All Posters
Poster A108 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Equivariant Self-Supervised Learning Improves IT Predictivity
Thomas Yerxa1 (), Jenelle Feather2, Eero Simoncelli1,2, SueYeon Chung1,2; 1Center for Neural Science, New York University, 2Center for Computational Neuroscience, Flatiron Institute
We present a novel method for self-supervised learning of representations that are equivariant to a set of transformations. When trained on images, we demonstrate that the learned representations effectively factorize sources of variability in their inputs, and provide improved prediction of responses of cells in macaque visual area IT across four different datasets.
Keywords: self-supervised learning factorized representations brain-model alignment