D-TRACE: Deep Triply-Aligned Clustering

Abstract

Deep clustering has recently emerged as a promising direction in clustering analysis, which aims to leverage the representation learning power of deep neural networks to enhance the clustering of highly-complex data. However, most of the existing deep clustering algorithms tend to utilize a single layer (typically the last fully-connected layer) of representation to build the clustering, yet cannot well exploit the rich and diverse information hidden in multiple layers. In view of this, this paper proposes a deep triply-aligned clustering (D-TRACE) approach, which is able to jointly explore three types of representations from multiple layers in the neural network. Specifically, we incorporate the contrastive learning into the first-stage network training, where three modules (i.e., the backbone network, the instance contrastive head, and the cluster contrastive head) are simultaneously optimized. By fusing the three types of representations from the three modules, we further propose the concept of the triply-aligned representation, based on which a unified neural network with a reconstruction loss and a Kullback-Leibler (KL) divergence based clustering loss is trained and thus the final clustering can be achieved in an unsupervised manner. Experiments on multiple image datasets demonstrate the superiority of our D-TRACE approach over the state-of-the-art deep clustering approaches.