The article introduces Consistent Assignment of Views over Random Partitions (CARP), a self-supervised clustering method for representation learning of visual features. The method optimizes a new pretext task based on random partitions of prototypes that regularizes the model and enforces consistency between views assignments. It improves training stability and prevents collapsed solutions in joint-embedding training. The paper shows that CARP’s representations are suitable for learning downstream tasks and outperforms many SSL methods in transfer learning tasks.
Publication date: 19 Oct 2023
Project Page: https://arxiv.org/abs/2310.12692v1
Paper: https://arxiv.org/pdf/2310.12692
