Representation learning for clustering via building consensus

Abstract In this paper, we focus on unsupervised representation learning for clustering of images. Recent advances in deep and are based the idea that different views an input image (generated through data augmentation techniques) must be close space (exemplar consistency), and/or similar images have cluster assignments (population consistency). We define additional notion consistency, consensus consistency , which ensures representations learned to induce partitions variations space, algorithms or initializations a single algorithm. loss by executing seamlessly integrate all three consistencies (consensus, exemplar population) into end-to-end framework. The proposed algorithm, using (ConCURL), improves upon performance state-of-the-art methods four out five datasets. Furthermore, extend evaluation procedure reflect challenges encountered real-world tasks, such as maintaining cases with distribution shifts. also perform detailed ablation study deeper understanding code trained models available at https://github.com/JayanthRR/ConCURL_NCE .