DualNet: Continual Learning, Fast and Slow

This project contains the implementation of the following NeurIPS 2021 paper:

Title: DualNet: Continual Learning, Fast and Slow (NeurIPS 2021), [openreview], [pdf].

Authors: Quang Pham, Chenghao Liu, and Steven Hoi

DualNet proposes a novel continual learning framework for (online) continual learning where a slow network gradually accumulates knowledge to improves its general representation over time and a fast network utilizes the slow representation to quickly learn new information.

Requirements

python 3.7.3
pytorch >= 1.8
torchvision >= 0.9
Kornia >= 0.5.0

Benchmarks

1. Prepare data

The data/ folders contains the train and test splits for the miniImageNet and CORE50 benchmarks. Download the raw data and modify the path in the csv files to point to the raw data folder.

2. Run experiments

To replicate our results on the task aware and task free settings, run

chmod +x scripts/*
./scripts/task_aware.sh
./scripts/task_free.sh

The results will be put in the resuts/ folders.

Semi-supervised learning setting

For the semi-supervised continual learning experiments, run

./scripts/semi.sh --rho 0.1
./scripts/semi.sh --rho 0.25

The parameter rho indicates the percentage of data that are labeled (e.g. 10% or 25%).

Citation

If you found DualNet useful for your research, please consider citing.

@article{pham2021dualnet,
  title={DualNet: Continual Learning, Fast and Slow},
  author={Pham, Quang and Liu, Chenghao and Hoi, Steven},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

phquang/DualNet