Reproduce “Few-Shot Incremental Learning with Continually Evolved Classifiers” on OpenResearch

About this paper

Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, Yinghui Xu

Zhang et al. introduce the Continually Evolved Classifier (CEC) framework for Few-Shot Class-Incremental Learning (FSCIL), which separates feature learning from classifier adaptation and employs a novel Pseudo Incremental Learning strategy. The approach improves the performance dropping rate (PD) by up to 18.83% on CUB200, enabling AI systems to learn new classes from limited examples while maintaining performance on existing knowledge.

Key insights

• Decoupling feature learning from classifier learning prevents catastrophic forgetting and overfitting when learning new classes from limited examples.
• A graph-based attention mechanism dynamically adapts classifier weights as new classes are added, maintaining global discriminability across all seen classes.
• Pseudo Incremental Learning (PIL) effectively trains the classifier adaptation module by simulating incremental scenarios using only base data, specifically leveraging image rotations to make familiar classes appear 'unfamiliar'.

Summary by alphaXiv.