From 3512592fa0a01638cc7007bd1822b9a94058178f Mon Sep 17 00:00:00 2001 From: "Anh M. Nguyen" Date: Fri, 12 Aug 2022 09:10:33 -0500 Subject: [PATCH] Added Visual correspondence-based explanations improve AI robustness and human-AI team accuracy --- README.md | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index ec2ec67..fa02cd6 100644 --- a/README.md +++ b/README.md @@ -232,17 +232,20 @@ This is an on-going attempt to consolidate interesting efforts in the area of un * L2E: Learning to Explain: Generating Stable Explanations Fast. _Situ et al. ACL 2021_ [pdf](https://aclanthology.org/2021.acl-long.415.pdf "Training neural networks to mimic a black-box attribution methods e.g. Occlusion, LIME, SHAP produces a faster and more stable explanation method.") | [code](https://github.com/situsnow/L2E) * Efficient Explanations from Empirical Explainers. _Schwarzenberg et al. 2021_ [pdf](https://arxiv.org/abs/2103.15429 "Training deep nets to approximate Integrated Gradient and Shapley methods") -### B2.3 Explaining by examples (prototypes) +### B2.3 Explaining by prototypes * This Looks Like That: Deep Learning for Interpretable Image Recognition. _Chen et al. NeurIPS 2019_ [pdf](https://arxiv.org/abs/1806.10574) | [code](https://github.com/cfchen-duke/ProtoPNet) * ProtoPNet: This Looks Like That, Because ... Explaining Prototypes for Interpretable Image Recognition. _Nauta et al. 2020_ [pdf](https://arxiv.org/pdf/2011.02863.pdf) - * NP-ProtoPNet: These do not Look Like Those. _Singh et al. 2021_ [pdf](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9373404 "ProtoPNet with negative prototypes and applied to chest x-rays") + * NP-ProtoPNet: These do not Look Like Those. _Singh et al. 2021_ [pdf](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9373404 "ProtoPNet with negative prototypes and applied to chest x-rays") + +### B2.4 Explaining by retrieving supporting examples +* Corr: Visual correspondence-based explanations improve AI robustness and human-AI team accuracy. _Nguyen, Taesiri, Nguyen 2022._ [pdf](https://arxiv.org/abs/2208.00780 "An interpretable-by-design XAI method that first retrieves similar patches (like kNN) to the input image from a training set or knowledgebase and then use them as evidence to label the input image. EMD-Corr and CHM-Corr improves OOD accuracy on ImageNet and improve human accuracy on CUB.") | [code](https://github.com/anguyen8/visual-correspondence-XAI) -### B2.4 Adversarial attacks on XAI systems with humans in the loop +### B2.5 Adversarial attacks on XAI systems with humans in the loop * When and How to Fool Explainable Models (and Humans) with Adversarial Examples. _Vadilo et al. 2021_ [pdf](https://arxiv.org/abs/2107.01943 "A framework of scenarios, assumptions, and humans in an XAI system under adversarial attacks") * The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. _Nguyen, Kim, Nguyen 2021_ [pdf](http://anhnguyen.me/project/feature-attribution-effectiveness/ "On image classification, feature attribution maps are less effective in improving human-AI team compared to a simple nearest-neighbor method. The effectiveness of heatmaps also does not correlate with their localization performance.") -### B2.5 Others +### B2.6 Others * Learning how to explain neural networks: PatternNet and PatternAttribution [pdf](https://arxiv.org/abs/1705.05598) * Deep Learning for Case-Based Reasoning through Prototypes [pdf](https://arxiv.org/pdf/1710.04806.pdf) * Unsupervised Learning of Neural Networks to Explain Neural Networks [pdf](https://arxiv.org/abs/1805.07468)