Awesome

BayLIME

BayLIME is a Bayesian modification of LIME (one of the most widely used approaches in XAI). Compared to LIME, BayLIME exploits prior knowledge and Bayesian reasoning to improve both the consistency in repeated explanations of a single prediction and the robustness to kernel settings. BayLIME also exhibits better explanation fidelity than the state-of-the-art (LIME, SHAP and GradCAM) by its ability to integrate prior knowledge from, e.g., a variety of other XAI techniques, as well as verification and validation (V&V) methods.

Publication

The paper on BayLIME is accepted by UAI2021, here is the accepted version on arXiv and a recorded presentation for UAI'21.

Setup

1. Copy-paste the modified_sklearn_BayesianRidge.py file (in the lime/utils folder on this repo) into your local sklearn.linear_model folder. To find out where the folder is, simply run:

from sklearn import linear_model
print(linear_model.__file__)

2. Download the necessary dataset for ImageNet and GTSRB model, unzip the files and move to the data folder.

ImageNet (original images): http://image-net.org/download-images
GTSRB (.h5 file): https://drive.google.com/file/d/1MjgsnH3bOYG_PvdvqmoamoCPmySQazRJ/view?usp=sharing

(Tested with Python version 3.7.3, scikit-learn version 0.22.1, Tensorflow version 2.0.0)

Repository Structure

• experiments contains the experiments of the paper, in which you may find both the code (in Python jupyter-notebook) and the original data generated (stored as HTML and .csv files).
• lime, all source-code of BayLIME that modifies the original LIME source-code can be found in this folder.
• data contains some data, e.g., images and tabular dataset.

Basic Use

The gist of BayLime is to allow users to embed informative prior knowledge when interpreting AI using Lime.

We have modified the Lime by adding more options to the args model_regressor.

Now when calling the explainer.explain_instance() API of BayLime, we have four options:

1. model_regressor='non_Bay' (default) uses sklearn Ridge regressor
2. model_regressor='Bay_non_info_prior' uses sklearn BayesianRidge regressor with all default args (fitting both hyperparameters alpha and lambda from samples)
3. model_regressor='Bay_info_prior' uses the modified sklearn BayesianRidge regressor and reads the hyperparameters alpha and lambda from configuration files,
4. model_regressor='BayesianRidge_inf_prior_fit_alpha' uses the modified BayesianRidge regressor and reads the hyperparameters lambda from configuration files and fit alpha from sampling data.

Please refer to the tutorials (e.g., BayLIME_tutorial_images.ipynb) for details.

Embed Prior from GradCAM

To get the explanation for a specific image (e.g. king penguin) in the data folder, firstly modify the image path in Line 46 of Grad_CAM_Prior.py, then type

python Grad_CAM_Prior.py

You will get the explanation results along with Deletion and Insertion AUC figures from GradCAM, LIME and BayLIME under the created evaluation_output folder.

To get statistical fidelity evaluation on ImageNet dataset, first please make sure the validation dataset from ImageNet called ILSVRC2012_img_val is already downloaded and moved to the data folder, then type

python del_ins_exp.py

You will get the explanation result for each image from ImageNet and a record file recording the runtime output in the created evaluation_output folder. Be cautious that the evaluation_output folder will be reset every time running the program, so take a copy if you want to save the results.

Embed Prior from Neural Cleanse

In backdoor_exp.py, we provide the explanations for backdoor inputs based on BadNet and TrojanAttack models. To get the IoU and AMD evaluations for the Prior, LIME and BayLIME, type the command

python backdoor_exp.py

You will get the print out of IoU and AMD scores for each backdoor attacked images. The interpretation of IoU and AMD scores can be referred to the paper.