Learning forward-compatible and domain-invariant representations for cross-domain few-shot class-incremental learning.

Abstract

Few-Shot Class-Incremental Learning (FSCIL) aims to continually recognize new classes from limited samples without forgetting previously learned ones, which typically assumes that all classes share the same domain. However, in real-world scenarios, incremental classes frequently originate from domains different from that of the base classes. Motivated by this, we introduce, for the first time, a more realistic and challenging problem: Cross-Domain Few-Shot Class-Incremental Learning (CDFSCIL), where the model is required to incrementally learn knowledge of new classes under both data scarcity and domain shift. To address this challenge, we propose a unified framework called Learning Forward-Compatible and Domain-Invariant Representations (FCDI). Through the joint design of three complementary components, FCDI learns representations that are compatible with future unseen classes and robust to domain shift. Specifically, FCDI constructs a structured representation space to preserve capacity for future classes, enabling novel categories to be incrementally integrated in a stable manner. To enhance domain robustness, semantic and domain features are explicitly disentangled so that the model focuses on learning domain-invariant semantic representations. In addition, FCDI exposes the model to diverse domain variations and encourages consistent representations across domains, which further promotes domain-invariant representation learning. Extensive experiments on both the proposed CDFSCIL benchmark and the standard FSCIL tasks demonstrate that our method significantly outperforms state-of-the-art approaches. The source code will be released.