Papers on Explainable Artificial Intelligence

This is an on-going attempt to consolidate interesting efforts in the area of understanding / interpreting / explaining / visualizing a pre-trained ML model.

---

GUI tools

• DeepVis: Deep Visualization Toolbox. Yosinski et al. ICML 2015 code | pdf
• SWAP: Generate adversarial poses of objects in a 3D space. Alcorn et al. CVPR 2019 code | pdf
• AllenNLP: Query online NLP models with user-provided inputs and observe explanations (Gradient, Integrated Gradient, SmoothGrad). Last accessed 03/2020 demo

Libraries

• CNN visualizations (feature visualization, PyTorch)
• iNNvestigate (attribution, Keras)
• DeepExplain (attribution, Keras)
• Lucid (feature visualization, attribution, Tensorflow)
• TorchRay (attribution, PyTorch)
• Captum (attribution, PyTorch)

Surveys

• Methods for Interpreting and Understanding Deep Neural Networks. Montavon et al. 2017 pdf
• Visualizations of Deep Neural Networks in Computer Vision: A Survey. Seifert et al. 2017 pdf
• How convolutional neural network see the world - A survey of convolutional neural network visualization methods. Qin et al. 2018 pdf
• A brief survey of visualization methods for deep learning models from the perspective of Explainable AI. Chalkiadakis 2018 pdf
• A Survey Of Methods For Explaining Black Box Models. Guidotti et al. 2018 pdf
• Understanding Neural Networks via Feature Visualization: A survey. Nguyen et al. 2019 pdf
• Explaining Explanations: An Overview of Interpretability of Machine Learning. Gilpin et al. 2019 pdf
• DARPA updates on the XAI program pdf
• Explainable Artificial Intelligence: a Systematic Review. Vilone at al. 2020 pdf

Opinions

• Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead Rudin et al. Nature 2019 pdf

Definitions of Interpretability

• The Mythos of Model Interpretability. Lipton 2016 pdf
• Towards A Rigorous Science of Interpretable Machine Learning. Doshi-Velez & Kim. 2017 pdf
• Interpretable machine learning: definitions, methods, and applications. Murdoch et al. 2019 pdf

Books

• A Guide for Making Black Box Models Explainable. Molnar 2019 pdf

A. Explaining inner-workings

A1. Visualizing Preferred Stimuli

Synthesizing images / Activation Maximization

• AM: Visualizing higher-layer features of a deep network. Erhan et al. 2009 pdf
• Deep inside convolutional networks: Visualising image classification models and saliency maps. Simonyan et al. 2013 pdf
• DeepVis: Understanding Neural Networks through Deep Visualization. Yosinski et al. ICML workshop 2015 pdf | url
• MFV: Multifaceted Feature Visualization: Uncovering the different types of features learned by each neuron in deep neural networks. Nguyen et al. ICML workshop 2016 pdf | code
• DGN-AM: Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Nguyen et al. NIPS 2016 pdf | code
• PPGN: Plug and Play Generative Networks. Nguyen et al. CVPR 2017 pdf | code
• Feature Visualization. Olah et al. 2017 url
• Diverse feature visualizations reveal invariances in early layers of deep neural networks. Cadena et al. 2018 pdf
• Computer Vision with a Single (Robust) Classifier. Santurkar et al. NeurIPS 2019 pdf | blog | code
• BigGAN-AM: Improving sample diversity of a pre-trained, class-conditional GAN by changing its class embeddings. Li et al. 2019 pdf

Real images / Segmentation Masks

• Visualizing and Understanding Recurrent Networks. Kaparthey et al. ICLR 2015 pdf
• Object Detectors Emerge in Deep Scene CNNs. Zhou et al. ICLR 2015 pdf
• Understanding Deep Architectures by Interpretable Visual Summaries. Godi et al. BMVC 2019 pdf

A2. Inverting Neural Networks

A2.1 Inverting Classifiers

• Understanding Deep Image Representations by Inverting Them. Mahendran & Vedaldi. CVPR 2015 pdf
• Inverting Visual Representations with Convolutional Networks. Dosovitskiy & Brox. CVPR 2016 pdf
• Neural network inversion beyond gradient descent. Wong & Kolter. NIPS workshop 2017 pdf

A2.2 Inverting Generators

• Image Processing Using Multi-Code GAN Prior. Gu et al. 2019 pdf

A3. Distilling DNNs into more interpretable models

• Interpreting CNNs via Decision Trees pdf
• Distilling a Neural Network Into a Soft Decision Tree pdf
• Distill-and-Compare: Auditing Black-Box Models Using Transparent Model Distillation. Tan et al. 2018 pdf
• Improving the Interpretability of Deep Neural Networks with Knowledge Distillation. Liu et al. 2018 pdf

A4. Quantitatively characterizing hidden features

• TCAV: Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors. Kim et al. 2018 pdf | code
  • Automating Interpretability: Discovering and Testing Visual Concepts Learned by Neural Networks. Ghorbani et al. 2019 pdf
• SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability. Raghu et al. 2017 pdf | code
• A Peek Into the Hidden Layers of a Convolutional Neural Network Through a Factorization Lens. Saini et al. 2018 pdf
• Network Dissection: Quantifying Interpretability of Deep Visual Representations. Bau et al. CVPR 2017 url | pdf
  • GAN Dissection: Visualizing and Understanding Generative Adversarial Networks. Bau et al. ICLR 2019 pdf
  • Net2Vec: Quantifying and Explaining how Concepts are Encoded by Filters in Deep Neural Networks. Fong & Vedaldi CVPR 2018 pdf
  • Intriguing generalization and simplicity of adversarially trained neural networks. Agarwal, Chen, Nguyen 2020 pdf
  • Understanding the Role of Individual Units in a Deep Neural Network. Bau et al. PNAS 2020 pdf

A5. Network surgery

• How Important Is a Neuron? Dhamdhere et al. 2018 pdf

A6. Sensitivity analysis

• NLIZE: A Perturbation-Driven Visual Interrogation Tool for Analyzing and Interpreting Natural Language Inference Models. Liu et al. 2018 pdf

B. Decision explanations

B1. Attribution maps

B1.0 Surveys

• Feature Removal Is A Unifying Principle For Model Explanation Methods. Covert et al. 2020 pdf

B1.1 White-box / Gradient-based

• A Taxonomy and Library for Visualizing Learned Features in Convolutional Neural Networks pdf

Gradient

• Deep inside convolutional networks: Visualising image classification models and saliency maps. Simonyan et al. 2013 pdf
• Deconvnet: Visualizing and understanding convolutional networks. Zeiler et al. 2014 pdf
• Guided-backprop: Striving for simplicity: The all convolutional net. Springenberg et al. 2015 pdf
• SmoothGrad: removing noise by adding noise. Smilkov et al. 2017 pdf

Input x Gradient

• DeepLIFT: Learning important features through propagating activation differences. Shrikumar et al. 2017 pdf
• Integrated Gradients: Axiomatic Attribution for Deep Networks. Sundararajan et al. 2018 pdf | code
  • Expected Gradients: Learning Explainable Models Using Attribution Priors. Erion et al. 2019 pdf | code
  • I-GOR: Visualizing Deep Networks by Optimizing with Integrated Gradients. Qi et al. 2019 pdf
  • BlurIG: Attribution in Scale and Space. Xu et al. CVPR 2020 pdf | code
  • XRAI: Better Attributions Through Regions. Kapishnikov et al. ICCV 2019 pdf | code
• LRP: Beyond saliency: understanding convolutional neural networks from saliency prediction on layer-wise relevance propagation pdf
  • DTD: Explaining NonLinear Classification Decisions With Deep Tayor Decomposition pdf

Activation map

• CAM: Learning Deep Features for Discriminative Localization. Zhou et al. 2016 code | web
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. Selvaraju et al. 2017 pdf
• Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. Chattopadhyay et al. 2017 pdf | code
• Smooth Grad-CAM++: An Enhanced Inference Level Visualization Technique for Deep Convolutional Neural Network Models. Omeiza et al. 2019 pdf
• NormGrad: There and Back Again: Revisiting Backpropagation Saliency Methods. Rebuffi et al. CVPR 2020 pdf | code

Learning the heatmap

• MP: Interpretable Explanations of Black Boxes by Meaningful Perturbation. Fong et al. 2017 pdf
  • MP-G: Removing input features via a generative model to explain their attributions to classifier's decisions. Agarwal et al. 2019 pdf | code
  • Understanding Deep Networks via Extremal Perturbations and Smooth Masks. Fong et al. ICCV 2019 pdf | code
• FIDO: Explaining image classifiers by counterfactual generation. Chang et al. ICLR 2019 pdf
• FG-Vis: Interpretable and Fine-Grained Visual Explanations for Convolutional Neural Networks. Wagner et al. CVPR 2019 pdf

Attributions of network biases

• Full-Gradient Representation for Neural Network Visualization. Srinivas et al. NeurIPS 2019 pdf
• Bias also matters: Bias attribution for deep neural network explanation. Wang et al. ICML 2019 pdf

Others

• Visual explanation by interpretation: Improving visual feedback capabilities of deep neural networks. Oramas et al. 2019 pdf
• Regional Multi-scale Approach for Visually Pleasing Explanations of Deep Neural Networks. Seo et al. 2018 pdfb

B1.2 Attention as Explanation

Computer Vision

• Multimodal explanations: Justifying decisions and pointing to the evidence. Park et al. CVPR 2018 pdf

NLP

• Attention is not Explanation. Jain & Wallace. NAACL 2019 pdf
• Attention is not not Explanation. Wiegreffe & Pinter. EMNLP 2019 pdf
• Learning to Deceive with Attention-Based Explanations. Pruthi et al. ACL 2020 pdf

B1.3 Black-box / Perturbation-based

• Sliding-Patch: Visualizing and understanding convolutional networks. Zeiler et al. 2014 pdf
• PDA: Visualizing deep neural network decisions: Prediction difference analysis. Zintgraf et al. ICLR 2017 pdf
• RISE: Randomized Input Sampling for Explanation of Black-box Models. Petsiuk et al. BMVC 2018 pdf
• LIME: Why should i trust you?: Explaining the predictions of any classifier. Ribeiro et al. 2016 pdf | blog
  • LIME-G: Removing input features via a generative model to explain their attributions to classifier's decisions. Agarwal et al. 2019 pdf | code
• SHAP: A Unified Approach to Interpreting Model Predictions. Lundberg et al. 2017 pdf | code
• OSFT: Interpreting Black Box Models via Hypothesis Testing. Burns et al. 2019 pdf

B1.4 Evaluating heatmaps

Computer Vision

• The (Un)reliability of saliency methods. Kindermans et al. 2018 pdf
• ROAR: A Benchmark for Interpretability Methods in Deep Neural Networks. Hooker et al. NeurIPS 2019 pdf | code
• Sanity Checks for Saliency Maps. Adebayo et al. 2018 pdf
• A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations. Nie et al. 2018 pdf
• BIM: Towards Quantitative Evaluation of Interpretability Methods with Ground Truth. Yang et al. 2019 pdf
• On the (In)fidelity and Sensitivity for Explanations. Yeh et al. 2019 pdf
• SAM: The Sensitivity of Attribution Methods to Hyperparameters. Bansal, Agarwal, Nguyen. CVPR 2020 pdf | code

NLP

• Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior? Hase & Bansal ACL 2020 pdf | code

B2. Learning to explain

B2.1 Regularizing attribution maps

• Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations. Ross et al. IJCAI 2017 pdf
• Learning Explainable Models Using Attribution Priors. Erion et al. 2019 pdf
• Interpretations are useful: penalizing explanations to align neural networks with prior knowledge. Rieger et al. 2019 pdf

B2.2 Explaining by examples (prototypes)

• This Looks Like That: Deep Learning for Interpretable Image Recognition. Chen et al. NeurIPS 2019 pdf | code
  • This Looks Like That, Because ... Explaining Prototypes for Interpretable Image Recognition. Nauta et al. 2020 pdf

B2.3 Others

• Learning how to explain neural networks: PatternNet and PatternAttribution pdf
• Deep Learning for Case-Based Reasoning through Prototypes pdf
• Unsupervised Learning of Neural Networks to Explain Neural Networks pdf
• Automated Rationale Generation: A Technique for Explainable AI and its Effects on Human Perceptions pdf
  • Rationalization: A Neural Machine Translation Approach to Generating Natural Language Explanations pdf
• Towards robust interpretability with self-explaining neural networks. Alvarez-Melis and Jaakola 2018 pdf

C. Counterfactual explanations

• Counterfactual Explanations for Machine Learning: A Review. Verma et al. 2020 pdf
• Interpreting Neural Network Judgments via Minimal, Stable, and Symbolic Corrections. Zhang et al. 2018 pdf
• Counterfactual Visual Explanations. Goyal et al. 2019 pdf
• Generative Counterfactual Introspection for Explainable Deep Learning. Liu et al. 2019 pdf

Generative models

• Generative causal explanations of black-box classifiers. O’Shaughnessy et al. 2020 pdf
• Removing input features via a generative model to explain their attributions to classifier's decisions. Agarwal et al. 2019 pdf | code

D. Others

• Yang, S. C. H., & Shafto, P. Explainable Artificial Intelligence via Bayesian Teaching. NIPS 2017 pdf
• Explainable AI for Designers: A Human-Centered Perspective on Mixed-Initiative Co-Creation pdf
• ICADx: Interpretable computer aided diagnosis of breast masses. Kim et al. 2018 pdf
• Neural Network Interpretation via Fine Grained Textual Summarization. Guo et al. 2018 pdf
• LS-Tree: Model Interpretation When the Data Are Linguistic. Chen et al. 2019 pdf