Variational Autoencoders Trained with Q-Deformed Lower Bounds

Abstract

Variational autoencoders (VAEs) have been successful at learning a low-dimensional manifold from high-dimensional data with complex dependencies. At their core, they consist of a powerful Bayesian probabilistic inference model, to capture the salient features of the data. In training, they exploit the power of variational inference, by optimizing a lower bound on the model evidence. The latent representation and the performance of VAEs are heavily influenced by the type of bound used as a cost function. Significant research work has been carried out into the development of tighter bounds than the original ELBO, to more accurately approximate the true log-likelihood. By leveraging the q-deformed logarithm in the traditional lower bounds, ELBO and IWAE, and the upper bound CUBO, we bring contributions to this direction of research. In this proof-of-concept study, we explore different ways of creating these q-deformed bounds that are tighter than the classical ones and we show improvements in the performance of such VAEs on the binarized MNIST dataset.