Learning Representation for Clustering via Dual Correlation

Contrastive learning is a self-supervised learning method used to learn the general features of a dataset without labels by letting the model learn which data points are similar or different. It constructs labeled samples through data augmentation. Therefore, the core and difficulty of self-supervised contrastive learning are to construct high-quality contrastive samples, especially negative sample pairs. In this paper, a cluster-based contrast sample selection method is proposed, which selects the positive contrast samples from the same cluster and the negative contrast samples from different clusters. By minimizing the forward contrast sample loss and maximizing the reverse sample loss, the deep representation clustering result of the image is obtained. The experimental results show the excellence and adaptability of the algorithm.