Twin Contrastive Learning for Online Clustering

This paper proposes to perform online clustering by conducting twin contrastive learning (TCL) at the instance and cluster level. Specifically, we find that when data is projected into a feature space with dimensionality of target number, rows columns its matrix correspond representation, respectively. Based on observation, for given dataset, proposed TCL first constructs positive negative pairs through augmentations. Thereafter, in row column matrix, instance- cluster-level are respectively conducted pulling together while pushing apart negatives. To alleviate influence intrinsic false-negative rectify assignments, adopt confidence-based criterion select pseudo-labels boosting both learning. As result, performance further improved. Besides elegant idea learning, another advantage it could independently predict assignment each instance, thus effortlessly fitting scenarios. Extensive experiments six widely-used image text benchmarks demonstrate effectiveness TCL. The code released https://pengxi.me .