Improving Variational Autoencoders Reconstruction Using Prototypes

Variational autoencoders implement latent space regularization with a known distribution, which enables stochastic synthesis from straightforward sampling procedures. Most of the various existing regularization approaches struggle with balancing proper regularization and good reconstruction quality. Here, we propose to circumvent such limitations by implementing a distributed latent space regularization using prototypical deep clustering with known distributions. Our approach enables high-quality data synthesis by simply decoding samples from such distributions around prototypes and disregards training any further auxiliary network to explicitly learn the decoder synthesis from codes. Such schema enables obtaining an appropriate number of clusters for solid regularization with better reconstruction quality and improved synthesis control. We experiment with our method using widespread exploratory benchmarks and report that regulariza-tion centered on optimal 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 coordinates.