ICCV 2025

Intervening in Black Box: Concept Bottleneck Model for Enhancing Human Neural Network Mutual Understanding (CBM-HNMU)

Paper

For getting the latest update of our paper, please refer to https://doi.org/10.48550/arXiv.2506.22803.

Citation

If you find this project helpful, please consider citing:

@InProceedings{Xiong_2025_ICCV,
    author    = {Xiong, Nuoye and Dong, Anqi and Wang, Ning and Hua, Cong and Zhu, Guangming and Mei, Lin and Shen, Peiyi and Zhang, Liang},
    title     = {Intervening in Black Box: Concept Bottleneck Model for Enhancing Human Neural Network Mutual Understanding},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {2836-2845}
}

requirement

clip==1.0  
jax==0.4.31  
jaxopt==0.8.3  
matplotlib==3.5.3  
numpy==1.19.5  
opencv_python==4.3.0.38  
Pillow==9.3.0  
Pillow==10.4.0  
scikit_learn==1.0.2  
scipy==1.7.3  
setuptools==65.6.3  
tensorflow_gpu==2.4.0  
timm==0.6.12  
torch==1.7.1  
torch_summary==1.4.5  
torchvision==0.8.2  

These are the main packages needed to be installed. For detailed, please refer to the requirement.txt.
We will also provide the integrated environment of Anaconda3 in the future.

Dependence

OpenAI-CLIP: https://github.com/openai/CLIP
CRAFT: https://github.com/deel-ai/Craft

Datasets

Please refer to the README.md in Dataset.

---

How to use it?

Train Baselines (Fine-Tune)

net switch : ["nfresnet50" , "vit" , "resnext26" , "botnet26t" , "rexnet100" , "gcvit" , "deit" , "convit" , "cait"]  
dataset switch : ["flower102" , "cifar10" , "cifar100" , "cub" , "aircraft"]  

python train_base.py <net_sw> <dataset_sw> <data_root>  
eg: python train_base.py nfresnet50 flower102 ../YOUR_FOLDER/Dataset  

Confusing Categories Selection

python Reasonable.py <net_sw> <dataset_sw> <data_root> <cc_select>  

Local Approximation

python CBM-HNMU.py <net_sw> <dataset_sw> <data_root> <ap> <tar_cls> <opt:ic_nums>  

Concepts Intervention

python CBM-HNMU.py <net_sw> <dataset_sw> <data_root> <ci> <tar_cls> <opt:ic_nums>  

Knowledge Transfer

python CBM-HNMU.py <net_sw> <dataset_sw> <data_root> <kt> <tar_cls> <opt:ic_nums>  

Visualization

python Reasonable.py <net_sw> <dataset_sw> <data_root> <reasonable> <tar_cls>