Unsupervised deep clustering via adaptive GMM modeling and optimization

Abstract Supervised deep learning techniques have achieved success in many computer vision tasks. However, most deep learning methods are data hungry and rely on a large number of labeled data in the training process. This work introduces an unsupervised deep clustering framework and studies the discovery of knowledge from a set of unlabeled data samples. Specifically, we propose a new network structure for both representation learning and GMM (Gaussian Mixture Model)-based representation modeling. In the training process of our proposed network, we not only adjust the Gaussian components to better model the distribution of representations, but also adjust the data representations towards their associating Gaussian centers to provide more adaptive support for the GMM. In this way, we take the data representations as the supervisory signal for the update of the GMM parameters and the GMM as the supervisory signal for the update of the representations, yet keeping the entire deep clustering as unsupervised. Consequently, we train the network based on an objective function with two learning targets. With the first target, we learn a GMM to model the representations properly and make each Gaussian component to be compact as much as possible. With the second target, we improve the inter-cluster distance by adapting the cluster centers to be further away from their neighbors. Thus, the training procedure simultaneously improves the intra-cluster compactness and inter-cluster separability for all the evolved clusters. Experimental results on eight datasets show that the proposed method can improve the clustering performance in comparison with the existing state of the art techniques.