Sparsity enforcement on latent variables for better disentanglement in VAE

A Study on the Latent Space of VAE by Inducing Sparsity in the Encoder Network

Abstract

We address the problem of unsupervised latent factorization and reconstruction accuracy. The related work on unsupervised representations focuses on constraining the second term of Variational Autoencoders loss function: The Kullback-Leibler component (Beta-VAE, FactorVAE Beta-TCVAE). Despite promising results, this comes with a trade-off between disentanglement and reconstruction. Besides, it is not clear why minimizing the KL divergence leads to disentanglement. In this paper, we propose to achieve disentangled representations by sampling from a sparse distribution. To give a visual appealing reconstruction for humans, we replace the conventional pixel-wise quadratic by perceptual loss. We demonstrate the reconstruction quality and disentangled on synthetic datasets.