Heterogeneous Tri-stream Clustering Network

Contrastive deep clustering has recently gained significant attention with its ability of joint contrastive learning and via neural networks. Despite the rapid progress, previous works mostly require both positive negative sample pairs for clustering, which rely on a relative large batch-size. Moreover, they typically adopt two-stream architecture two augmented views, overlook possibility potential benefits multi-stream architectures (especially heterogeneous or hybrid networks). In light this, this paper presents new end-to-end approach termed Heterogeneous Tri-stream Clustering Network (HTCN). The tri-stream in HTCN consists three main components, including weight-sharing online networks target network, where parameters network are exponential moving average that Notably, trained by simultaneously (i) predicting instance representations (ii) enforcing consistency between cluster Experimental results four challenging image datasets demonstrate superiority over state-of-the-art approaches. code is available at https://github.com/dengxiaozhi/HTCN .