This repository is the official implementation of Homeostasis-Inspired Continual Learning: Learning to Control Structural Regularization.

To install requirements:

pip install -r requirements.txt

It is recommended to use the Anaconda

To train the Homeostatic Meta-Model (HM) in the paper, run this command:

python meta_alpha_train.py --data MNISTBPERM --n_task 30 --seed 20

or Execute the pre-established shell script:

./run.sh

To evaluate Homeostatic Meta-Model on MNIST-BPERM, run:

python meta_alpha_test.py --data MNISTPERM

Using homeostatic meta-trained model, you can evaluate the performance on continual learning.

You can also evaluate alternative models for comparison

• Single : a single learner based on SGD for a sequence of tasks
• InDep : a dedicated (independent) learner based on SGD for each task
• EWC : Elastic Weight Consolidation (regularized with the dedicated Fisher information for each task) [1]
• OEWC : Online Elastic Weight Consolidation (regularized with the accumulated Fisher information) [2]
• IMM : Incremental Moment Matching with a weight transfer method [3]
• Multi : Multi-task learning (allowed to access all the tasks, violation of strict CL scenario)

python train.py --model InDep --data MNISTBPERM --n_task 30 --seed 0 

Use "--model" argument with above model names to train other alternatives

Our model achieves the following performance on the sequence of 10 MNIST-PERM tasks:

[1] Kirkpatrick, James, et al. "Overcoming catastrophic forgetting in neural networks." Proceedings of the national academy of sciences 114.13 (2017): 3521-3526.

[2] Schwarz, Jonathan, et al. "Progress & compress: A scalable framework for continual learning." arXiv preprint arXiv:1805.06370 (2018).

[3] Lee, Sang-Woo, et al. "Overcoming catastrophic forgetting by incremental moment matching." Advances in neural information processing systems. 2017.