Awesome

Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning

This is the code repo for Multi-layer Rehearsal Feature Augmentation for Class-Incremental Learning (ICML 2024).

[paper]

TL; DR: We analyze the overfitting for rehearsal samples in class-incremental learning (CIL) from the perspective of all-layer margin. We find that such margin is getting smaller and smaller during the training of CIL. We propose the Multi-layer Rehearsal Feature Augmentation (MRFA) to enlarge the margin of the rehearsal samples during CIL.

To Run the Code

python main.py --config {config_file_path}

Acknowledgments

This repo uses PyCIL.