Unsupervised Disentanglement Learning via Dirichlet Variational Autoencoder

Unsupervised disentanglement learning is the process of discovering factorized variables that include interpretable semantic information and encode separate factors of variations in the data. It is a critical learning problem and has been applied in various tasks and domains. Most of the existing unsupervised disentanglement learning methods are based on the variational autoencoder (VAE) and adopt Gaussian distribution as the prior over the latent space. However, these methods suffer from a collapse of the decoder weights, which leads to degraded disentangling ability, due to the Gaussian prior. To address this issue, in this paper we propose a novel unsupervised disentanglement learning method based on a VAE framework in which the Dirichlet distribution is deployed as the prior over latent space. In our method, the interpretable factorised latent representations can be obtained by balancing the capacity of the latent information channel and the learning of statistically independent latent factors. The effectiveness of our method is validated through experiments on several publicly available datasets.