Prototypical Variational Autoencoders

Dario Augusto Borges Oliveira, Laura Elena Cue La Rosa

Keywords: Variational Autoencoders, Latent Space Regularization

Abstract: Variational autoencoders are unsupervised generative models that implement latent space regularization towards a known distribution, enabling stochastic synthesis from straightforward sampling procedures. Many works propose various regularization approaches, but most struggle to compromise between proper regularization and good reconstruction quality. This paper proposes distributing the regularization through the latent space using prototypical anchored clusters, each with an optimal position in the latent space and following a known distribution. Such schema enables obtaining an appropriate number of clusters with solid regularization for better reconstruction quality and improved synthesis control. We experiment with our method using widespread exploratory benchmarks and report that regularization anchored on prototypes' coordinates or cluster centroids neutralizes the adverse effects regularization terms often have on autoencoder reconstruction quality, matching non-regularized autoencoders' performance. We also report appealing results for interpreting data representatives with simple prototype synthesis and controlling the synthesis of samples with prototype-like characteristics from decoding white noise around prototype anchors.

One-sentence Summary: A new method for variational auto encoders regularization using prototypical online clustering