Overcoming Catastrophic Forgetting in Incremental Few-Shot Learning by Finding Flat Minima

This paper considers incremental few-shot learning, which requires a model to\ncontinually recognize new categories with only a few examples provided. Our\nstudy shows that existing methods severely suffer from catastrophic forgetting,\na well-known problem in incremental learning, which is aggravated due to data\nscarcity and imbalance in the few-shot setting. Our analysis further suggests\nthat to prevent catastrophic forgetting, actions need to be taken in the\nprimitive stage -- the training of base classes instead of later few-shot\nlearning sessions. Therefore, we propose to search for flat local minima of the\nbase training objective function and then fine-tune the model parameters within\nthe flat region on new tasks. In this way, the model can efficiently learn new\nclasses while preserving the old ones. Comprehensive experimental results\ndemonstrate that our approach outperforms all prior state-of-the-art methods\nand is very close to the approximate upper bound. The source code is available\nat https://github.com/moukamisama/F2M.\n