This repository contains code and scripts for the following paper:
Concept Bottleneck Models
Pang Wei Koh*, Thao Nguyen*, Yew Siang Tang*, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang
ICML 2020
The experiments use the following datasets:
To download the TravelingBirds dataset, which we use to test robustness to background shifts, please download the CUB_fixed folder from this CodaLab bundle by clicking on the download button. If you use this dataset, please also cite the original CUB and Places datasets.
The NIH Osteoarthritis Initiative (OAI) dataset requires an application for data access, so we are unable to provide the raw data here. To access that data, please first obtain data access permission from the Osteoarthritis Initiative, and then refer to this Github repository for data processing code. If you use it, please cite the Pierson et al. paper corresponding to that repository as well.
Here, we focus on scripts replicating our results on CUB, which is public. We provide an executable, Dockerized version of those experiments on CodaLab.
We seek to learn models that we can interact with using high-level concepts: would the model predict severe arthritis if it thinks there is a bone spur in the x-ray? State-of-the-art models today do not typically support the manipulation of concepts like "the existence of bone spurs", as they are trained end-to-end to go directly from raw input (e.g., pixels) to output (e.g., arthritis severity). We revisit the classic idea of first predicting concepts that are provided at training time, and then using these concepts to predict the label. By construction, we can intervene on these concept bottleneck models by editing their predicted concept values and propagating these changes to the final prediction. On x-ray grading and bird identification, concept bottleneck models achieve competitive accuracy with standard end-to-end models, while enabling interpretation in terms of high-level clinical concepts ("bone spurs") or bird attributes ("wing color"). These models also allow for richer human-model interaction: accuracy improves significantly if we can correct model mistakes on concepts at test time.
We used the same environment as Codalab's default gpu setting, please run pip install -r requirements.txt. Main packages are:
- matplotlib 3.1.1
- numpy 1.17.1
- pandas 0.25.1
- Pillow 6.2.2
- scipy 1.3.1
- scikit-learn 0.21.3
- torch 1.1.0
- torchvision 0.4.0
Note that we updated Pillow and removed tensorflow-gpu and tensorboard from requirements.txt.
You can pull the Docker image directly from Docker Hub.
docker pull codalab/default-gpu
Standard task training for CUB can be ran using the scripts/experiments.sh and Codalab scripts can be ran using scripts/codalab_experiments.sh. More information about how to perform data processing and other evaluations can be found in the README in CUB/.