Evidential Sparsification of Multimodal Latent Spaces in Conditional Variational Autoencoders

We consider the problem of sparsifying the discrete latent space of a trained conditional variational autoencoder, while preserving its learned multimodality.