Concept-Driven Continual Learning

This is the official implemenation of our TMLR 2024 paper: "Concept-Driven Continual Learning".

• In this work, we propose to bring interpretability into continual learning process to mitigate catastrophic forgetting problems.
• We proposed two novel methods Interpretability-Guided Continual Learning (IG-CL) and Intrinsically-Interpretable Neural Network (IN2) that can systematically manage human-understandable concepts within DNNs throughout the training process.
• Our proposed approaches provide unprecedented transparency and control over the continual learning process, marking a promising new direction in designing continual learning algorithms.

Overview of our Method 1: IG-CL

Overview of our Method 2: IN2

Setup

Execute the following code to set up the code environment.

bash setup/setup.sh

For methods related to MIR and DER, please follow setup/new_env.md to modify the environment and code.

Running IG-CL

1. Run and evaluate IG-CL strategy

Execute the following code to get the experiment results. Here we take CIFAR-10 for example.

python continual_learning/train_all.py --result_dir results --strategy SRT --task_num 5 --dataset cifar10 --sol sol0

sol0 means freeze-all implementation, and sol1 means freeze-part implementation.

2. Get experiment results

Execute the following code to get the experiment results.

python evaluate/metric.py --file_dir results --strategy SRT --task_num 5 

Running IN2

1. Get labels and classes for each task

• We need to know the classes in each task to create the corresponding concept set.
• Follow steps in continual_learning/label_generation.ipynb to get the classes for each task.

2. Create Concept Set

Execute the following code to generate the concept set. Please put your OpenAI api key in sandbox-lf-cbm/.openai_api_key before running the code.

bash script_dir/CBM/exec_conceptset.sh

• For specific scenario, please modify the variables seed_list and task_num in sandbox-lf-cbm/GPT_conceptset_processor.py and sandbox-lf-cbm/GPT_init_concepts.py.
• Please see Label-free Concept Bottleneck Models for more details.

3. Run and evaluate IN2 strategy

Execute the following code to train a model using IN2 strategy.

bash script_dir/CBM/cc_cbm.sh

4. Get experiment results

Execute the following code to get the experiment results.

python evaluate/metric.py --file_dir results/cc_cbm --strategy cc_cbm --task_num 5 

More experiment results

• Please check experiments/ folder for experiments in the paper
• Forward Transfer Metric: Use avalanche.evaluation.metrics.forward_transfer metric to get the results.

Sources

• CLIP-Dissect: https://github.com/Trustworthy-ML-Lab/CLIP-dissect
• Label-free CBM: https://github.com/Trustworthy-ML-Lab/Label-free-CBM
• Avalanche: https://avalanche.continualai.org

Citation

@article{yang2024conceptdriven,
  title={Concept-Driven Continual Learning},
  author={Yang, Sin-Han and Oikarinen, Tuomas and Weng, Tsui-Wei},
  journal={Transactions on Machine Learning Research},
  year={2024}
}