PDF

Interpreting and explaining the behavior of neural networks is critical for many tasks. Explainable AI provides a way to address this challenge, mostly by providing per-pixel importance on the decision in a post-hoc paradigm. Yet, interpreting such explanations may still require expert knowledge. Some recent attempts toward interpretability adopt a concept-based framework, giving a higher-level relationship between some concepts and model decisions. This paper proposes Bottleneck Concept Learner (BotCL). BotCL represents an image solely by the presence/absence of concepts learned through training over the target task without explicit supervision over the concepts. BotCL uses tailored regularizers so that learned concepts can be human-understandable. Using some image classification tasks as our testbed, we demonstrate BotCL's potential to remodel neural networks for better interpretability.

Download CUB or ImageNet and set them into direction of your "dataset_dir". You can also make your own dataset with the structure similar to ImageNet and name it as Custom.

Using this command to get number and labels of concepts

python conceptnet.py 

Using the following command for training

python main_recon.py --num_classes 10 --num_cpt 20 --lr 0.001 --epoch 50 --lr_drop 30

Use the following command for the inference of a sample. You can change the index to select different input samples. Change top-sample (top-10 in the paper) to show more details for a concept. Change deactivate to remove a concept. Visualization for the input sample and all concepts are shown at folder "vis" and "vis_pp", respectively.

python vis_recon.py --num_classes 10 --num_cpt 20 --index 0 --top_sample 20 ---deactivate -1

We first pre-train the backbone and then train the whole model. For ImageNet and Custom, just change the name for dataset.

Pre-training of backbone:
python main_contrast.py --num_classes 50 --num_cpt 20 --base_model resnet18 --lr 0.0001 --epoch 60 --lr_drop 40 --pre_train True --dataset CUB200 --dataset_dir "your dir"

Training for BotCL:
python main_contrast.py --num_classes 50 --num_cpt 20 --base_model resnet18 --lr 0.0001 --epoch 60 --lr_drop 40 --dataset CUB200 --dataset_dir "your dir" --weak_supervision_bias 0.1 --quantity_bias 0.1 --distinctiveness_bias 0.05 --consistence_bias 0.01

Use the following commend to visualize the learned concept.

First run process.py to extarct the activation for all dataset samples:
python process.py

Then see the generated concepts by:
python vis_contrast.py --num_classes 50 --num_cpt 20 --base_model resnet18 --index 100 --top_sample 20 --dataset CUB200

If you want to use this work, please consider citing the following paper.

@inproceedings{wang2023botcl,
 author = {Bowen Wang and Liangzhi Li and Yuta Nakashima and Hajime Nagahara},
 booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
 pages = {},
 title = {Learning Bottleneck Concepts in Image Classification},
 year = {2023}
}