Neighborhood Commonality-aware Evolution Network for Continuous Generalized Category Discovery

TL;DR

NCENet tackles Continuous Generalized Category Discovery (C-GCD) by decomposing the problem into learning discriminative representations for novel classes (NCRL) and preserving old-class knowledge (BCKD) in an incremental, unlabelled-data setting. NCRL leverages local neighborhood commonalities via μ_i = (1/k) sum_{q in NN(z_i)} z_q to form meaningful prediction distributions p over μ, guided by a self-distillation objective with L_ncrl = -1/|B| sum_i sum_j p_i^j log \, p_hat_i^j, enabling robust novel-class representation without requiring explicit labels. BCKD introduces a bi-level contrastive distillation framework with student-anchored and teacher-anchored losses L_sa and L_ta, and sets L_bckd = (L_sa + L_ta)/2 to balance plasticity and stability during incremental learning. Experiments on CIFAR10, CIFAR100 and Tiny-ImageNet show NCENet achieving state-of-the-art clustering accuracy, especially in the last incremental sessions, validating the effectiveness of neighborhood-based representation learning and contrastive knowledge transfer for continual generalized category discovery.

Abstract

Continuous Generalized Category Discovery (C-GCD) aims to continually discover novel classes from unlabelled image sets while maintaining performance on old classes. In this paper, we propose a novel learning framework, dubbed Neighborhood Commonality-aware Evolution Network (NCENet) that conquers this task from the perspective of representation learning. Concretely, to learn discriminative representations for novel classes, a Neighborhood Commonality-aware Representation Learning (NCRL) is designed, which exploits local commonalities derived neighborhoods to guide the learning of representational differences between instances of different classes. To maintain the representation ability for old classes, a Bi-level Contrastive Knowledge Distillation (BCKD) module is designed, which leverages contrastive learning to perceive the learning and learned knowledge and conducts knowledge distillation. Extensive experiments conducted on CIFAR10, CIFAR100, and Tiny-ImageNet demonstrate the superior performance of NCENet compared to the previous state-of-the-art method. Particularly, in the last incremental learning session on CIFAR100, the clustering accuracy of NCENet outperforms the second-best method by a margin of 3.09\% on old classes and by a margin of 6.32\% on new classes. Our code will be publicly available at \href{https://github.com/xjtuYW/NCENet.git}{https://github.com/xjtuYW/NCENet.git}. \end{abstract}

Figures (7)

• Figure 1: Each class consists of a set of local commonalities that are shared between instances within the same neighborhoods, our proposed NCRL exploits prediction distributions over these local commonalities to guide the learning of representational differences between instances of different classes.
• Figure 2: Pipeline of our proposed incremental learning framework. Our proposed method leverages the Neighborhood Commonality-aware Representation Learning (NCRL) module to learn representations for novel classes and the Bi-level Contrastive Knowledge Distillation (BCKD) module to maintain the representation ability for old classes. In NCRL, local commonalities $\mu$ derived from neighborhoods are used to generate prediction distributions, and a self-distillation technique is used to learn representations. In BCKD, student-anchored contrastive knowledge distillation and teacher-anchored contrastive knowledge distillation are performed to achieve holistic representational knowledge retention.
• Figure 3: Clustering accuracy in each incremental learning session under different balance factor ${\lambda}_b$. Our proposed method prefers a small balance factor.
• Figure 4: Clustering accuracy in the last incremental session under different temperature ${\tau}$.
• Figure 5: Clustering accuracy in the last incremental learning session under different numbers of selected neighbors. A relatively larger number can help our proposed method achieve better performance.