Selective Concept Bottlenecks Without Predefined Concepts (TMLR 2025)

Official code for our paper "Selective Concept Bottlenecks Without Predefined Concepts" (TMLR 2025).

If you find this work useful, please consider citing our work:

@article{schrodi2025selective,
    title={Selective Concept Bottlenecks Without Predefined Concepts},
    author={Schrodi, Simon and Schur, Julian and Argus, Max and Brox, Thomas},
    journal={Transactions on Machine Learning Research},
    year={2025}
}

Here is an overview of our method, UCBM:

Setup

1. Install the conda environment, via conda env create --name ucbm --file env.yml

2. call conda develop .

3. set the base path in constants.py (this is where everything will be saved) and the other paths if necessary.

4. Download the models following the instructions from Trustworthy-ML-Lab/Label-free-CBM and adjust the paths in constants.py if necessary.

5. Download the CUB dataset with bash download_cub.sh

Training of UCBMs

Step 1: Discover concepts

ImageNet:

python discover_concepts.py -d imagenet -b resnet_v2 --con_am 3

CUB:

python discover_concepts.py -d cub -b cub_rn18 --con_am 1

Places-365:

python discover_concepts.py -d places365 -b places365_rn18 --con_am 5

Step 2: Train interpretable classifier

ImageNet:

python train_cbm.py -d imagenet -b resnet_v2 --concept_data "concepts_3000_64" --epochs 20 --lam_gate 0 --lam_w 1e-4 --dropout_p 0.1 --lr 0.001 --cls_save_name "topk_seed_0" --scale_choose 'no' --bias_choose 'learn' --normalize_concepts --relu 'no' --k 42 --seed 0

CUB:

python train_cbm.py -d cub -b cub_rn18 --concept_data "concepts_200_64" --epochs 20 --lam_gate 0 --lam_w 8e-4 --dropout_p 0.2 --lr 0.001 --cls_save_name "topk_seed_0" --scale_choose 'no' --bias_choose 'learn' --normalize_concepts --relu 'no' --k 66 --seed 0

Places-365:

python train_cbm.py -d places365 -b places365_rn18 --concept_data "concepts_1825_64" --epochs 20 --lam_gate 0 --lam_w 4e-4 --dropout_p 0.2 --lr 0.01 --cls_save_name "topk_seed_0" --scale_choose 'no' --bias_choose 'learn' --normalize_concepts --relu 'no' --k 162 --seed 0

Visualization of explanations

cbm_name is the defined cls_save_name from above with a timestamp.

Plot explanations of individual decisions:

python explanation/prediction.py -d imagenet -b resnet50_v2 -c concepts_3000_64 --cbm_name $cbm_name

Plot explanations of class:

python explanation/class.py -d imagenet -b resnet50_v2 -c concepts_3000_64 --cbm_name $cbm_name

Acknowledgements

We thank the following GitHub users for their contributions which are used in this repository:

• CRAFT from deel-ai/Craft
• Base cpu NMF implemenation scikit-learn/scikit-learn