Papers about explanation

NIPS 2019

1. Grid Saliency for Context Explanations of Semantic Segmentation

2. Deliberative Explanations: visualizing network insecurities

3. On the (In)fidelity and Sensitivity for Explanations
   麻蛋，太难了看不懂

4. Explanations can be manipulated and geometry is to blame

5. CXPlain: Causal Explanations for Model Interpretation under Uncertainty

6. GNNExplainer: Generating Explanations for Graph Neural Networks

7. Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain)

8. ✔️Saccader: Accurate, Interpretable Image Classification with Hard Attentionf

9. ✔️Fooling Neural Network Interpretations via Adversarial Model Manipulation
   这篇文章和AAAI的1，以及Arvix的1有很大的相似性。
   本文训练了网络的参数，使得保持目标图像的类别不变的情况下，interpretability解释发生很大的变化。比如，向图像边缘变化，曾经的topk不再是topk等等。

10. Learning Dynamics of Attention: Human Prior for Interpretable Machine Reasoning

11. A Benchmark for Interpretability Methods in Deep Neural Networks

12.✔️ Towards Automatic Concept-based Explanations

13. ✔️ This Looks Like That: Deep Learning for Interpretable Image Recognition

ICML 2019

1. Bias Also Matters: Bias Attribution for Deep Neural Network Explanation
   use bias to generate explanation

2. Explaining Deep Neural Networks with a Polynomial Time Algorithm for Shapley Value Approximation

3. Counterfactual Visual Explanations

4. Interpreting Adversarially Trained Convolutional Neural Networks
   用smoothed backpropagation 解释了一种adversariallly trained network的robustness

5. On the Connection Between Adversarial Robustness and Saliency Map Interpretability

ICCV 2019

ICLR2019

1.✔️ LEARNING WHAT AND WHERE TO ATTEND
提出一个click maps游戏，让人标记显著的区域，然后设计网络，提取soft attention层，训练的时候设计loss，让soft attention层和人类标记的重要区域相似。

2. ✔️Explaining Image Classifiers by Counterfactual Generation
   针对Interpolation的方法的两个问题进行改进：
   － 利用droupout改进生成很多mask耗时的问题。
   － 利用GAN改进遮挡图片时用的不是环境信息（distribution改变）的问题。

3.✔️ VISUAL EXPLANATION BY INTERPRETATION: IMPROVING VISUAL FEEDBACK CAPABILITIES OF DEEP NEURAL NETWORKS]
generate visual explanation same as CAM
create a fake flower dataset.

4. ✔️APPROXIMATING CNNS WITH BAG-OF-LOCAL FEATURES MODELS WORKS SURPRISINGLY WELL
   ON IMAGENET

CVPR2019

1. Learning to Explain with Complemental Examples
   ICML的文章Learning to Explain: An Information-Theoretic Perspective on Model Interpretation是这篇文章的基础，不过我看不懂啊啊啊啊啊啊啊啊啊啊。。。。

2. Explainable and Explicit Visual Reasoning over Scene Graphs

3. Attention Branch Network:
   Learning of Attention Mechanism for Visual Explanation

4. End-to-end Interpretable Neural Motion Planner

5. Interpretable and Fine-Grained Visual Explanations for
   Convolutional Neural Networks

AAAI2019

1. ✔️Interpretation of Neural Networks is Fragile
   use adversarial examples to attack the network. make the network output unchanged but the attention map changed

Arxiv

1. ✔️Interpretable Deep Learning under Fire

ICML 2018

1. Learning to Explain: An Information-Theoretic Perspective
   on Model Interpretation

2. A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations

3. Discovering Interpretable Representations for Both Deep Generative and Discriminative Models

4. Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)

ICLR2018

1. LEARNING HOW TO EXPLAIN NEURAL NETWORKS: PATTERNNET AND PATTERNATTRIBUTION

2. Revealing interpretable object representations from human behavior

3. Hierarchical interpretations for neural network predictions

NIPS2017

1. Real Time Image Saliency for Black Box Classifiers
   类似Ruth Fong的方法，和RISE/LIME是一个类型的，都是image perturbation。
   用到了adversarial training的思路。使mask的面积尽量小，而且尽量的平滑。

   
   
   

   训练过程没有看