Official implementation of “CI-CBM: Class-Incremental Concept Bottleneck Model for Interpretable Continual Learning“.

Published at Transactions on Machine Learning Research (TMLR).

CI-CBM is an interpretable class-incremental learning method that combines concept regularization with pseudo-concept generation, enabling both local and global explanations for decision-making throughout class-incremental phases. It maintains robust concept-based reasoning as new classes are introduced, substantially outperforms prior interpretable continual-learning methods, and remains competitive with black-box baselines in both pretrained and non-pretrained settings.

Model overview and visual results

Setup

To set up the environment and download necessary datasets and models, follow the instructions below.

1. Install Python (Version 3.8 is recommended.)
2. Install dependencies by running pip install -r requirements.txt
3. Install CLIP and SigLip: pip install git+https://github.com/openai/CLIP.git and pip install open_clip_torch
4. Download the datasets by following files in the dataset_download folder:
   • CUB: bash download_cub.sh
   • ImageNet Subset: bash download_imagenetsubset.sh and imagenet_val_grouping.py
   • TinyImageNet: bash download_tinyimagenet and group the validation samples by their class using tinyimagenet_val_grouping.py
   • ImageNet: bash download_imagenet.sh and group the validation samples by their class using imagenet_val_grouping.py
   • Set dataset paths: After downloading, open data_utils.py and replace YOUR_PATH_TO_DATASET with the path to your dataset directory (e.g., the folder containing CUB, ImageNet, tiny-imagenet-200, etc.).
5. Download pretrained backbones (as needed for your experiments):
   • From the project root, run:

     cd backbone_download
     bash download_rn18_places.sh
     bash download_rn18_FeTrIL.sh
     bash download_apg.sh

   • Each script moves to the project root, creates backbone_checkpoints/, downloads the files there, then returns to the project root.

Running the models

1. Train CI-CBM

Train a Class-Incremental Concept Bottleneck Model by running main.py (see training_commands.md for examples).

Important parameters:

• seed: Random seed (default: 1993)
• backbone: One of the backbones listed above
• strategy: backbone_prototype (default), naive, or full_rehearsal
• clip_name: CLIP or SigLIP model for concept alignment (e.g. ViT-B/16, ViT-L-16-SigLIP-384)
• dataset: cifar10, cifar100, cub, tiny_imagenet, imagenetsubset, places365
• SAGA_lr: Learning rate for final layer (default: 0.1)
• n_iters: Iterations for final layer (default: 1000)
• lam: Sparsity regularization (default: 0.0007)
• n_experiences: Number of incremental phases
• half_split: Use large first phase; required for first-phase trained backbones

2. Evaluate trained models

Evaluate the trained models by running evaluate_cbm.py codes in the training_commands.md to calculate metrics.

Citation

If you use the FeTrIL or APG backbones in your research, please cite the following works:

FeTrIL (GitHub):

@inproceedings{petit2023fetril,
  title={Fetril: Feature translation for exemplar-free class-incremental learning},
  author={Petit, Gr{\'e}goire and Popescu, Adrian and Schindler, Hugo and Picard, David and Delezoide, Bertrand},
  booktitle={Proceedings of the IEEE/CVF winter conference on applications of computer vision},
  pages={3911--3920},
  year={2023}
}

APG (GitHub):

@inproceedings{tang2023prompt,
  title={When prompt-based incremental learning does not meet strong pretraining},
  author={Tang, Yu-Ming and Peng, Yi-Xing and Zheng, Wei-Shi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={1706--1716},
  year={2023}
}

SAGA (sparse final layer training) (GitHub):

@inproceedings{wong2021leveraging,
  title={Leveraging sparse linear layers for debuggable deep networks},
  author={Wong, Eric and Santurkar, Shibani and Madry, Aleksander},
  booktitle={International Conference on Machine Learning},
  pages={11205--11216},
  year={2021},
  organization={PMLR}
}