Reproduce “Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks” on OpenResearch

About this paper

Da-Wei Zhou, Han-Jia Ye, Liang Ma, Di Xie, Shiliang Pu, De-Chuan Zhan

A meta-learning framework named LIMIT leverages a Transformer-based calibration module to enable models to learn new classes from few examples incrementally without forgetting old knowledge. It demonstrates improved performance, outperforming prior state-of-the-art by 2.09% in last session accuracy and reducing the performance dropping rate by 1.35% on CIFAR100.

Key insights

• Simulating the multi-phase incremental learning process during base training, rather than a single-stage extension, is crucial for building a generalizable feature space for long-term adaptation.
• A Transformer-based set-to-set function effectively bridges the semantic gap between old and new class representations by refining embeddings and providing instance-specific contextualization.
• The framework simultaneously mitigates catastrophic forgetting of old classes and reduces overfitting to limited new class data by jointly learning feature embeddings and a calibration strategy.

Summary by alphaXiv.