Deep Fuzzy Clustering with Weighted Intra-class Variance and Extended Mutual Information Regularization

Recently, many joint deep clustering methods, which simultaneously learn latent embedding and predict clustering assignments through deep neural network, have received a lot of attention. Among these methods, KL divergence based clustering framework is one of the most popular branches. However, the clustering performances of these methods depend on an additional auxiliary target distribution. In this paper, we build a novel deep fuzzy clustering (DFC) network to learn discriminative and balanced assignment without the need of any auxiliary distribution. Specifically, we design an elaborate fuzzy clustering layer (FCL) to estimate more discriminative assignments, and utilize weighted intra-class variance (WIV) as clustering objective function to enhance the compactness of the learned embedding. Moreover, we propose extended mutual information (EMI) between input data and the corresponding clustering assignments as a regularization to achieve "fair" but "firm" assignment. Extensive experiments conducted on several datasets illustrate the superiority of the proposed approach comparing to the state-of-the-art methods.