Deep Adversarial Gaussian Mixture Auto-Encoder for Clustering

Warith Harchaoui, Pierre-Alexandre Mattei, Charles Bouveyron

Abstract: Feature representation for clustering purposes consists in building an explicit or implicit mapping of the input space onto a feature space that is easier to cluster or classify. This paper relies upon an adversarial auto-encoder as a means of building a code space of low dimensionality suitable for clustering. We impose a tunable Gaussian mixture prior over that space allowing for a simultaneous optimization scheme. We arrive at competitive unsupervised classification results on hand-written digits images (MNIST) that is customarily classified within a supervised framework.

TL;DR: We simultaneously train both an auto-encoder and a Gaussian mixture model for clustering in an adversarial fashion

Conflicts: uw.edu, parisdescartes.fr

Keywords: Deep learning, Unsupervised Learning