Learning to Cluster with Auxiliary Tasks: A Semi-Supervised Approach

In this paper, we propose a model to learn a feature-category latent representation of the data, that is guided by a semi-supervised auxiliary task. The goal of this auxiliary task is to assign labels to unlabeled data and regularize the feature space. Our model is represented by a modified version of a Categorical Variational Autoencoder, i.e., a probabilistic generative model that approximates a categorical distribution with variational inference. We benefit from the autoencoder's architecture to learn powerful representations with Deep Neural Networks in an unsupervised way, and to optimize the model with semi-supervised tasks. We derived a loss function that integrates the probabilistic model with our auxiliary task to guide the learning process. Experimental results show the effectiveness of our method achieving more than 90% of clustering accuracy by using only 100 labeled examples. Moreover we show that the learned features have discriminative properties that can be used for classification.