A Note on Deep Variational Models for Unsupervised Clustering

Rui Shu, James Brofos, Curtis Langlotz

Abstract: Recently, the Gaussian Mixture Variational Autoencoder (GMVAE) has been introduced to handle unsupervised clustering (Dilokthanakul et al., 2016). However, the existing formulation requires the introduction of the free bits term into the objective function in order to overcome the effects of the uniform prior imposed on the latent categorical variable. By considering our choice of generative and inference models, we propose a simple variation on the GMVAE that performs well empirically without modifying the variational objective function.

TL;DR: We make small changes to the generative and inference models and show huge impact on clustering/classification.

Keywords: Deep learning, Unsupervised Learning

Conflicts: stanford.edu, mitre.org