Continual Novel Class Discovery via Feature Enhancement and Adaptation

TL;DR

This work tackles Continual Novel Class Discovery (CNCD) by addressing feature-discrepancy and inter-session confusion through a novel Feature Enhancement and Adaptation (FEA) framework. FEA employs a single-head, prior-distribution-guided guide-to-novel framework, augmented by a Centroid-to-Samples Similarity Constraint (CSS) and a Boundary-Aware Prototype Constraint (BAP) to both enrich feature diversity and stabilize prototype space across incremental sessions. The approach abandons auxiliary cluster heads and exemplar storage, instead leveraging contrastive learning, prior distributions, and prototype-aware losses to progressively discover novel classes while preserving known ones. Empirical results on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate state-of-the-art performance, especially in long-incremental protocols up to ten sessions, with thorough ablations validating the contributions of CSS and BAP to robustness and discriminability.

Abstract

Continual Novel Class Discovery (CNCD) aims to continually discover novel classes without labels while maintaining the recognition capability for previously learned classes. The main challenges faced by CNCD include the feature-discrepancy problem, the inter-session confusion problem, etc. In this paper, we propose a novel Feature Enhancement and Adaptation method for the CNCD to tackle the above challenges, which consists of a guide-to-novel framework, a centroid-to-samples similarity constraint (CSS), and a boundary-aware prototype constraint (BAP). More specifically, the guide-to-novel framework is established to continually discover novel classes under the guidance of prior distribution. Afterward, the CSS is designed to constrain the relationship between centroid-to-samples similarities of different classes, thereby enhancing the distinctiveness of features among novel classes. Finally, the BAP is proposed to keep novel class features aware of the positions of other class prototypes during incremental sessions, and better adapt novel class features to the shared feature space. Experimental results on three benchmark datasets demonstrate the superiority of our method, especially in more challenging protocols with more incremental sessions.

Figures (4)

• Figure 1: Comparison of the multi-head paradigm (GM, FRoST, etc.) and our single-head paradigm to tackle the feature-discrepancy problem. (a) The multi-head paradigm requires pseudo-labels from an auxiliary cluster head as self-supervision signals. (b) Our single-head paradigm uses prior distribution as a self-supervision signal. (c) Accuracy comparison of different methods on Tiny-ImageNet with 10 incremental discovery sessions. The left figure illustrates the accuracy of known classes at the end of each incremental session, and the right figure shows the average accuracy of novel classes across all sessions. ‘S1’ represents the 1-st incremental session.
• Figure 2: An overview of our FEA for CNCD. The proposed FEA is comprised of a guide-to-novel framework, a centroid-to-samples similarity constraint (CSS), and a boundary-aware prototype constraint (BAP). The guide-to-novel framework can continually discover novel knowledge while maintaining old knowledge without any auxiliary cluster head. The CSS is designed to enhance the distinctiveness between novel class features. The BAP is proposed to better adapt novel class features to the shared feature space.
• Figure 3: The overall accuracy of all learned classes (i.e., known classes) at the end of each incremental discovery session on CIFAR-10, CIFAR-100, and Tiny-ImageNet, where T indicates the number of incremental discovery sessions.
• Figure 4: The comparison of confusion matrices on Tiny-ImageNet under T=10 incremental protocol. The warmer color represents the higher accuracy.