Learn the Time to Learn: Replay Scheduling in Continual Learning

Marcus Klasson, Hedvig Kjellström, Cheng Zhang

Accepted to Transactions on Machine Learning Research (TMLR) in 09/2023.

This repository includes Pytorch code for all experiments in the paper.

Openreview: https://openreview.net/forum?id=Q4aAITDgdP

Arxiv: https://arxiv.org/abs/2209.08660

Installation

• Create conda environment from file:

conda env create -f environment.yml
conda activate rs_in_cl

• The following structure is in the main directory:

/data                        : Directory for preprocessing datasets not provided by torchvision
/mcts_single_cl_envs         : Directory with code for MCTS experiments
/rl_policy_generalization    : Directory with code for DQN and A2C experiments

Entering the directories mcts_single_cl_envs/ and rl_policy_generalization/ to see code used in Section 4.1 and 4.2 respectively.

Citation

If you find this work useful in your research, please cite our paper:

@article{klasson2023learn,
    title={Learn the Time to Learn: Replay Scheduling in Continual Learning},
    author={Marcus Klasson and Hedvig Kjellstr{\"o}m and Cheng Zhang},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2023},
    url={https://openreview.net/forum?id=Q4aAITDgdP},
    note={}
}