Adaptive exploration for few-shot incremental learning

Abstract

Few-shot class incremental learning (FSCIL) poses a challenging problem in computer vision, where conventional deep models suffer from catastrophic forgetting and overfitting to novel classes. Inspired by the dynamic learning processes observed in human cognition when adapting to unfamiliar scenarios, we propose a deep exploratory incremental learning framework that incrementally refines the classifier model through a trial-and-error decision making process. A joint distribution-aware reward function is introduced to guide learning, incorporating three key factors: intra-class compactness, inter-class dispersion, and cross-session consistency, enabling balanced knowledge retention and acquisition. Furthermore, we design a dynamic gradient guidance module that adaptively adjusts gradient updates within a Gaussian-derived policy space, enhancing training stability and mitigating overfitting risks in the few shot regime. Extensive experiments conducted on three publicly available datasets demonstrate the effectiveness of the proposed method, achieving state-of-the-art performance in the FSCIL setting.