Few-shot Class-Incremental Learning via Generative Co-Memory Regularization

This work introduces a generative co-memory regularization approach for Few-shot Class-Incremental Learning (FSCIL). The method leverages generative domain adaptation to fine-tune a pre-trained encoder on a few base class examples, incorporating a masked autoencoder decoder and a classifier for efficient representation learning. It employs class-wise representation and weight memories to collaboratively regularize incremental learning, improving recognition accuracy while mitigating catastrophic forgetting and overfitting.