Efficient Representation Learning for Generalized Category Discovery

Zilong Xu, Chao Tan, Jiaxi Zhang, Anning Song, Genlin Ji

Abstract: Highlights•Addressed key issues: limited mini-batch samples and false positives in GCD.•Proposed a robust Divide-and-Combine strategy for efficient representation learning.•Analyzed the cluster-level alignment and uniformity of contrastive loss in GCD.•Unified system for efficient representation learning and reliable category discovery.