phquang/continual-normalization

Continual Normalization: Rethinking Batch Normalization for Online Continual Learning

This project contains the implementation of the following ICLR 2022 paper:

Title: Continual Normalization: Rethinking Batch Normalization for Online Continual Learning (ICLR 2022), [openreview]

Authors: Quang Pham, Chenghao Liu, and Steven Hoi

Overview

Continual Normalization (CN) is a simple, yet effective normalization strategy specially deveoped for the online continual learning problem. CN is highly compatible with state-of-the-art experience replay based methods and offers improvements over the traditional Batch Normalization strategy.

Usage

CN is simple and easy to implement. A standalone implementation of CN is provded in the cn.py file.

Replacing BN with CN in existing models

In many experiments, it is more convenient to consider a pre-built or pre-trained models and replace its BN layer with our CN, while keeping the pre-trained affine transformation parameters. We provide an utility function to do so and a working example in the example.py file.

Replicating the Online Class Incremental Learning Experiments

Lastly, to replicate the Online Class Incremental Learning Experiments, please follow the instructions in the mammoth/ folder.

Citing CN

If you found our work to be useful, please consider citing as

@inproceedings{pham2021continual,
  title={Continual Normalization: Rethinking Batch Normalization for Online Continual Learning},
  author={Pham, Quang and Liu, Chenghao and Steven, HOI},
  booktitle={International Conference on Learning Representations},
  year={2022}
}