Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style Julius von Kügelgen 1, 2 Y ash Sharma 3, 4 Luigi Gresele 1 Wieland Brendel

Neural Information Processing Systems 

Unlike prior work on disentanglement and independent component analysis, we allow for both nontrivial statistical and causal dependencies in the latent space.

Similar Docs  Excel Report  more

TitleSimilaritySource
None found