A PyTorch implementation of the continual learning experiments with deep neural networks described in the following paper:

• Brain-inspired replay for continual learning with artificial neural networks: https://www.nature.com/articles/s41467-020-17866-2

This paper proposes a new, brain-inspired version of generative replay that can scale to continual learning problems with natural images as inputs. This is demonstrated with the Split CIFAR-100 protocol, both for task-incremental learning and for class-incremental learning.

The current version of the code has been tested with Python 3.5.2 on several Linux operating systems with the following versions of PyTorch and Torchvision:

• pytorch 1.1.0
• torchvision 0.2.2

The versions that were used for other Python-packages are listed in requirements.txt.

To use the code, download the repository and change into it:

git clone https://github.com/GMvandeVen/brain-inspired-replay.git
cd brain-inspired-replay

(If downloading the zip-file, extract the files and change into the extracted folder.)

Assuming Python and pip are set up, the Python-packages used by this code can be installed using:

pip install -r requirements.txt

However, you might want to install pytorch and torchvision in a slightly different way to ensure compatability with your version of CUDA (see https://pytorch.org/).

Finally, the code in this repository itself does not need to be installed, but a number of scripts should be made executable:

chmod  x main_*.py compare_*.py create_figures.sh

./main_cl.py --experiment=splitMNIST --scenario=class --replay=generative --brain-inspired --pdf

This runs a single continual learning experiment: brain-inspired replay on the class-incremental learning scenario of split MNIST. Information about the data, the model, the training progress and the produced outputs (e.g., a pdf with results) is printed to the screen. Expected run-time on a standard laptop is ~12 minutes, with a GPU it should take ~4 minutes.

./compare_MNIST.py --scenario=class

This runs a series of continual learning experiments to compare the performance of various methods. Information about the different experiments, their progress and the produced outputs (e.g., a summary pdf) is printed to the screen. Expected run-time on a standard laptop is ~50 minutes, with a GPU it should take ~18 minutes.

These two demos can also be run with on-the-fly plots using the flag --visdom. For this visdom must be activated first, see instructions below.

The script create_figures.sh provides step-by-step instructions for re-running the experiments and re-creating the figures reported in the paper.

Although it is possible to run this script as it is, it will take very long and it is probably sensible to parallellize the experiments.

Using main_cl.py, it is possible to run custom individual experiments. The main options for this script are:

• --experiment: which task protocol? (splitMNIST|permMNIST|CIFAR100)
• --scenario: according to which scenario? (task|domain|class)
• --tasks: how many tasks?

To run specific methods, use the following:

• Context-dependent-Gating (XdG): ./main_cl.py --xdg --xdg-prop=0.8
• Elastic Weight Consolidation (EWC): ./main_cl.py --ewc --lambda=5000
• Online EWC: ./main_cl.py --ewc --online --lambda=5000 --gamma=1
• Synaptic Intelligenc (SI): ./main_cl.py --si --c=0.1
• Learning without Forgetting (LwF): ./main_cl.py --replay=current --distill
• Generative Replay (GR): ./main_cl.py --replay=generative
• Brain-Inspired Replay (BI-R): ./main_cl.py --replay=generative --brain-inspired

For information on further options: ./main_cl.py -h.

PyTorch-implementations for several methods relying on stored data (Experience Replay, iCaRL and A-GEM), as well as for additional metrics (FWT, BWT, forgetting, intransigence), can be found here: https://github.com/GMvandeVen/continual-learning.

With this code it is possible to track progress during training with on-the-fly plots. This feature requires visdom. Before running the experiments, the visdom server should be started from the command line:

python -m visdom.server

The visdom server is now alive and can be accessed at http://localhost:8097 in your browser (the plots will appear there). The flag --visdom should then be added when calling ./main_cl.py to run the experiments with on-the-fly plots.

For more information on visdom see https://github.com/facebookresearch/visdom.

Please consider citing our paper if you use this code in your research:

@article{vandeven2020brain,
  title={Brain-inspired replay for continual learning with artificial neural networks},
  author={van de Ven, Gido M and Siegelmann, Hava T and Tolias, Andreas S},
  journal={Nature Communications},
  volume={11},
  pages={4069},
  year={2020}
}

The research project from which this code originated has been supported by an IBRO-ISN Research Fellowship, by the Lifelong Learning Machines (L2M) program of the Defence Advanced Research Projects Agency (DARPA) via contract number HR0011-18-2-0025 and by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DoI/IBC) contract number D16PC00003. Disclaimer: views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of DARPA, IARPA, DoI/IBC, or the U.S. Government.