Improving Latent Representations via Explicit Disentanglement

A promising approach for improving both the interpretability and usefulness of latent representations for downstream tasks is disentanglement. Though recent work in variational frameworks has found success implicitly encouraging disentanglement for learning better representations, such methods do not take advantage of the readily available priors in most data. We propose three methods for explicit disentanglement and evaluate their ability to learn better representations with different datasets. Additionally, we investigate how we can leverage explicit disentanglement for learning representations under biased data.