Efficient Representation Learning for Generalized Category Discovery

Abstract

Generalized Category Discovery (GCD) aims to discover and cluster data from previously unseen classes. Numerous studies have delved into GCD; however, they either overlooked the significance of representation learning or naively adapt the contrastive loss from self-supervised learning without further refinements. Therefore, in this paper, we propose a simple yet effective method to address this task. Our first contribution is to design a robust Divide-and-Combine strategy with an improved contrastive loss to generate informative features. Subsequently, in light of the scarce supervised signals, we propose a Neighborhood-based Sampling module based on the analysis of cluster-level alignment and uniformity to better discover novel categories. Ultimately, to boost the model’s ability to learn from unlabeled data, we adopt a fusion training approach utilizing combinational patch features and the deterministic threshold to selectively identify high-confidence samples. We unify these learning processes within a single framework and jointly optimize. Extensive experiments across three general datasets and four fine-grained benchmark datasets show the compelling evidence for the superiority of our proposed method—Efficient Representation Learning for Generalized Category Discovery (ERL4GCD).