CFASL: Composite Factor-Aligned Symmetry Learning for Disentanglement in Variational Autoencoder

Hee-Jun Jung, Jaehyoung Jeong, Kangil Kim

Keywords: Disentanglement learning, Symmetries, Variational AutoEncoder

Abstract:

Implemented symmetries of input and latent vectors is important for disentanglement learning in VAEs, but most works focus on disentangling each factor without consideration of multi-factor change close to real world transformation between two samples, and even a few studies to handle it in autoencoder literature are constrained to pre-defined factors. We propose a novel disentanglement framework
for Composite Factor-Aligned Symmetry Learning (CFASL) on VAEs for the extension to general multi-factor change condition without constraint. CFASL disentangles representations by 1) aligning their changes, explicit symmetries, and unknown factors via proposed inductive bias, 2) building a composite symmetry for multi-factor change between two samples, and 3) inducing group equivariant
encoder and decoder in the condition. To set up the multi-factor change condition, we propose sample pairing for inputs, and an extended evaluation metric. In quantitative and in-depth qualitative analysis, CFASL shows significant improvement of disentanglement in multi-factor change condition compared to state-of-the-art methods and also gradually improves in single factor change condition on common benchmarks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2152