Mutual-Taught Deep Clustering

Deep clustering seeks to group data into distinct clusters using deep learning techniques. Existing approaches of deep clustering can be broadly categorized into two groups: offline clustering based on unsupervised representation learning and online clustering based on unsupervised classification. While both groups have demonstrated impressive performance in deep clustering, no study has explored the integration of their respective strengths. To this end, we propose Mutual-Taught Deep Clustering (MTDC), which unifies unsupervised representation learning and unsupervised classification into a framework while realizing mutual promotion using a novel mutual-taught mechanism. Specifically, MTDC alternates between predicting pseudolabels in label space and estimating semantic similarity in feature space during training. Moreover, pseudolabels provide weakly-supervised information to enhance unsupervised representation learning, while semantic similarities function as structural priors that regularize unsupervised classification. Consequently, unsupervised classification and unsupervised representation learning can mutually benefit from one another. MTDC is decoupled from prevailing deep clustering methods. For the sake of clarity, we build upon a straightforward baseline in this paper. Despite its simplicity, we demonstrate that MTDC is exceedingly efficacious and consistently enhances the baseline results by substantial margins. For example, MTDC achieves 2.5% similar to 7.9% (NMI), 3.0% 13.9% (ACC), and 3.1% similar to 16.7% (ARI) gains over the baseline on six widely used image datasets.

Automated summary

This paper proposes a method called Mutual-Taught Deep Clustering (MTDC) that integrates unsupervised representation learning and unsupervised classification. By alternating between predicting pseudolabels and estimating semantic similarity during training, MTDC allows unsupervised classification and unsupervised representation learning to mutually benefit from each other. Experimental results show that this method performs well on multiple image datasets.