Learning optimal conditional priors for disentangled representations

A large part of the literature on learning disentangled representations focuses on variational autoencoders (VAEs). Recent developments demonstrate that disentanglement cannot be obtained in a fully unsupervised setting without inductive biases on models and data. As such, Khemakhem et al. (2020) suggest employing a factorized prior distribution over the latent variables that is conditionally dependent on auxiliary observed variables complementing input observations. While this is a remarkable advancement toward model identifiability, the learned conditional prior only focuses on sufficiency, giving no guarantees on a minimal representation. Motivated by information theoretic principles, we propose a novel VAE-based generative model with theoretical guarantees on disentanglement. Our proposed model learns a sufficient and compact – thus optimal – conditional prior, which serves as regularization for the latent space. Experimental results indicate superior performance with respect to state-of-the-art methods, according to several established metrics proposed in the literature on disentanglement.