Variance Fingerprints: Component-Wise Identification of Content Factors by Augmentation Design

D. Hudson Smith

Augmentation-based self-supervised learning can recover the content shared across views without separating its underlying factors into individual coordinates. We ask whether augmen…