Label-free Concept Bottleneck Models

This is the official repository for our paper Label-Free Concept Bottleneck Models published at ICLR 2023. LF-CBM is a new framework to transform any neural networks into an interpretable Concept Bottleneck Model (CBM) without the need for labeled concept data.

Update 6/5/2023: We have conducted a crowdsourced user study to measure the interpretability of our method, results are available in Appendix B of our arxiv paper.

Setup

1. Install Python (3.9) and PyTorch (1.13).
2. Install dependencies by running pip install -r requirements.txt
3. Download pretrained models by running bash download_models.sh (they will be unpacked to saved_models)
4. Download and process CUB dataset by running bash download_cub.sh
5. Download ResNet18(Places365) backbone by running bash download_rn18_places.sh

We do not provide download instructions for ImageNet data, to evaluate using your own copy of ImageNet you must set the correct path in DATASET_ROOTS["imagenet_train"] and DATASET_ROOTS["imagenet_val"] variables in data_utils.py.

Running the models

1. Creating Concept Sets (Optional):

A. Create initial concept set using GPT-3 - GPT_initial_concepts.ipynb, do this for all 3 prompt types (can be skipped if using the concept sets we have provided). NOTE: This step costs money and you will have to provide your own openai.api_key.

B. Process and filter the conceptset by running GPT_conceptset_processor.ipynb (Alternatively get ConceptNet concepts by running ConceptNet_conceptset.ipynb)

2. Train LF-CBM

Train a concept bottleneck model on CIFAR10 by running:

python train_cbm.py --concept_set data/concept_sets/cifar10_filtered.txt

3. Evaluate trained models

Evaluate the trained models by running evaluate_cbm.ipynb. This measures model accuracy, creates barplots explaining individual decisions and prints final layer weights which are the basis for creating weight visualizations.

Additional evaluations and reproductions of our model editing experiments are available in the notebooks of experiments directory.

Results

High Accuracy:

			Dataset
Model	CIFAR10	CIFAR100	CUB200	Places365	ImageNet
Standard	88.80%	70.10%	76.70%	48.56%	76.13%
Standard (sparse)	82.96%	58.34%	75.96%	38.46%	74.35%
Label-free CBM	86.37%	65.27%	74.59%	43.71%	71.98%

For commands to train Label-free CBM and Standard (sparse) models on all 5 datasets, see training_commands.txt.

Explainable Decsisions:

Sources

CUB dataset: https://www.vision.caltech.edu/datasets/cub_200_2011/

Sparse final layer training: https://github.com/MadryLab/glm_saga

Explanation bar plots adapted from: https://github.com/slundberg/shap

CLIP: https://github.com/openai/CLIP

Cite this work

T. Oikarinen, S. Das, L. Nguyen and T.-W. Weng, Label-free Concept Bottleneck Models, ICLR 2023.

@inproceedings{oikarinenlabel,
  title={Label-free Concept Bottleneck Models},
  author={Oikarinen, Tuomas and Das, Subhro and Nguyen, Lam M and Weng, Tsui-Wei},
  booktitle={International Conference on Learning Representations},
  year={2023}
}