RED Server Software Copyright (C) 2022 Cognizant Digital Business, Evolutionary AI. All Rights Reserved.

red-paper

Code and supporting materials for the AAAI 2022 RED paper

This repository contains all the source codes to reproduce the experimental results reported in paper "Detecting Misclassification Errors in Neural Networks with a Gaussian Process Model", which is published in AAAI 2022. (Arxiv Link: https://arxiv.org/abs/2010.02065)

Steps to reproduce the main results in the paper:

1. Before running the source code, download the 121 UCI datasets from https://github.com/bioinf-jku/SNNs, and the 4 newly added datasets from UCI Machine Learning Repository, namely, "Phishing","messidor","Bioconcentration", and "Climate".
2. Put the 121 UCI datasets into a folder named UCI121_data, and the 4 newly added dataset into a folder named Datasets.
3. Create a folder named Results to store all the experimental results.
4. Create a folder named Statistics to store all the analysis results.
5. pip install the packages in requirements.txt.
6. Run main_experiments_UCI.py to generate all the experimental results.
7. Run analysis_results_UCI.py to analyze the stored experimental results.

Note: the packages used in calculating Trust Score can be git cloned from https://github.com/google/TrustScore.

Steps to reproduce other experimental results in the paper:

1. For results of "SVGP" variant, run experiments_SVGP.py after obtaining the main results in Results folder.
2. For results of "DNGO" variant, run experiments_DNGO.py after obtaining the main results in Results folder. The source codes of DNGO can be git cloned from https://github.com/automl/pybnn/blob/master/pybnn/dngo.py.
3. For results of "Entropy" variant, it can be directly calculated from the outputs of original NN classifier using entropy function from scipy.stats.
4. For results regarding Bayesian Neural Networks, run experiments_BNN.py to generate the results of base BNN classifier, then run experiments_BNN+RED.py to get the results of applying RED on top of BNN classifiers.
5. For results regarding MC-dropout, run experiments_dropout.py to generate the results of base NN classifier with MC-dropout, then run experiments_dropout+RED.py to get the results of applying RED on top of NN classifiers with MC-dropout.
6. For results regarding VGG16/VGG19 on CIFAR-10/CIFAR-100, train the base VGG16/VGG19 classifiers using code from https://github.com/geifmany/cifar-vgg, then run the included codes that generate results on UCI datasets on CIFAR-10/CIFAR-100 results.
7. For results regarding OOD and adversarial samples, run experiments_OOD_adversarial.py.
8. For results regarding OOD detection in CIFAR-10 vs. SVHN, directly apply the trained VGG16 and RED models in step 6 to SVHN data downloaded from http://ufldl.stanford.edu/housenumbers/.
9. For results of BLR-residual, run experiments_BLR-residual.py after running main_experiments_UCI.py.

Citation

If you use RED in your research, please cite it using the following BibTeX entry

@article{qiu:aaai22, 
         title={Detecting Misclassification Errors in Neural Networks with a Gaussian Process Model}, 
         volume={36}, 
         url={https://ojs.aaai.org/index.php/AAAI/article/view/20773}, 
         DOI={10.1609/aaai.v36i7.20773}, 
         number={7}, 
         journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
         author={Qiu, Xin and Miikkulainen, Risto}, 
         year={2022}, 
         month={Jun.}, 
         pages={8017-8027} }