All-about-XAI

This repository is all about papers and tools of Explainable AI

Contents

• Surveys

• Visualization Systems/Tools

• XAI Methods:

• 1. Transparent Models

• 2. Post-Hoc Explainability

  • Model-Agnostic

    • Visualization Method
    • Model Distillation
    • Adverserial Attack
    • Feature Relevance/Importance Method
    • Local interpretability

  • Model-Specific

    • Tree-based
    • CNN
      • Visualization
      • Explainable Model
      • Model Modification
    • RNN
    • Generative Model
    • Reinforcement learning

Surveys

Towards better analysis of machine learning models: A visual analytics perspective.Liu, S., Wang, X., Liu, M., & Zhu, J. (2017). Visual Informatics, 1, 48-56.

Visual Interpretability for Deep Learning: a Survey Quanshi Zhang, Song-Chun Zhu (2018) CVPR

interpretable/disentangled middle-layer representations

Towards a rigorous science of interpretable machine learning. F. Doshi-Velez and B. Kim. (2018).

Trends and trajectories for explainable, accountable and intelligible systems: An HCI research agenda. A. Abdul, J. Vermeulen, D. Wang, B. Y. Lim, and M. Kankanhalli,in Proc. SIGCHI Conf. Hum. FactorsComput. Syst. (CHI), 2018, p. 582

most focus on HCI research

A survey of methods for explaining black box models. R. Guidotti, A. Monreale, F. Turini, D. Pedreschi, and F. Giannotti.(2018).

presented a detailed taxonomy of explainability methods according to the type of problem faced.

Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI) A. Adadi and M. Berrada,in IEEE Access, vol. 6, pp. 52138-52160, 2018.

Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez.arxiv.(2019)

How convolutional neural network see the world - A survey of convolutional neural network visualization methods. Qin, Z., Yu, F., Liu, C., & Chen, X. (2018). ArXiv, abs/1804.11191.

Visualization Systems/Tools

• explAIner

  explAIner: A Visual Analytics Framework for Interactive and Explainable Machine Learning, Spinner, T., Schlegel, U., Schäfer, H., & El-Assady, M. (2019). IEEE VAST, Transactions on Visualization and Computer Graphics, 26, 1064-1074.

• CSI: collaborative semantic inference

  Visual Interaction with Deep Learning Models through Collaborative Semantic Inference. Gehrmann, S., Strobelt, H., Krüger, R., Pfister, H., & Rush, A.M. (2019). IEEE VAST. Transactions on Visualization and Computer Graphics, 26, 884-894.

  User can both understand and control parts of the model reasoning process. eg. in text summarization system, user can collaborative writing a summary with machines suggestion.

• Manifold

  Manifold: A Model-Agnostic Framework for Interpretation and Diagnosis of Machine Learning Models.Zhang, J., Wang, Y., Molino, P., Li, L., & Ebert, D.S. (2019). IEEE VAST, Transactions on Visualization and Computer Graphics, 25, 364-373.

  inspection (hypothesis), explanation (reasoning), and refinement (verification)

  

• ActiVis

  Activis: Visual exploration of industry-scale deep neural network models. M. Kahng, P. Y. Andrews, A. Kalro, and D. H. P. Chau. IEEE transactions on visualization and computer graphics, 24(1):88–97, 2018 Facebook

  compares the activations from different data instances (i.e., examples) to investigate the potential causes of misclassifications.

  

XAI Method

Transparent Model

As long as the model is accurate for the task, and uses a reasonably restricted number of internal components, intrinsic interpretable models are suffcient. Otherwise, use post-hoc methods.

Decision Trees

General Additive Models

Bayesian Models

Post-Hoc Explainability

Including natural language explanations, visualizations of learned models , and explanations by example.

Model Agnostic

Visualization Method

1. Saliency

Interpretable explanations of black boxes by meaningful perturbation, R. C. Fong, A. Vedaldi, in IEEE International Conference on Computer Vision, 2017, pp. 3429–3437.

Real time image saliency for black box classifiers, P. Dabkowski, Y. Gal, in: Advances in Neural Information Processing Systems, 2017, pp. 6967–6976.

2. Sensitivity

Sensitivity refers to how an ANN output is influenced by its input and/or weight perturbations

Opening black box data mining models using sensitivity analysis, P.Cortez and M.J.Embrechts, in Proc. IEEE Symp.Comput.Intell.Data Mining (CIDM), (2011)

Using sensitivity analysis and visualization techniques to open black box data mining models, P. Cortez and M. J. Embrechts,Inf. Sci. (2013).

3. SHAP: SHapley Additive exPlanations

Assign importance values for each feature, for a given prediction based on the game theoretic concept of Shapley values

A unified approach to interpreting model predictions, S.M. Lundberg and S.I. Lee, in Proc. Adv. Neural Inf. Process. Syst., 2017.

4. Partial Dependence Plot (PDP)

• PDP

Auditing black-box models for indirect influence, P. Adler, C. Falk, S. A. Friedler, T. Nix, G. Rybeck, C. Scheidegger, B. Smith, S. Venkatasubramanian, Knowledge and Information Systems (2018)

• ICE: Individual Conditional Expectation(extends PDP)

Peeking inside the black box: Visualizing statistical learning with plots of individual conditional expectation, A. Goldstein, A. Kapelner, J. Bleich, E. Pitkin, Journal of Computational and Graphical Statistics 24 (1) (2015) 44–65.

ICE plots extend PDP, reveal interactions and individual differences by disaggregating the PDP output.

• PI & ICI

Visualizing the feature importance for black box models, G. Casalicchio, C. Molnar, B. Bischl, Joint European Conference on Machine Learning and Knowledge Discovery in Databases,Springer, 2018, pp. 655–670

5. Surrogate Models

LIME

Interpretability via model extraction. O. Bastani, C. Kim, and H. Bastani. (2017).

TreeView: Peeking into deep neural networks via feature-space partitioning. J. J. Thiagarajan, B. Kailkhura, P. Sattigeri, and K. N. Ramamurthy.(2016)

6. Loss Function Vis

Visualizing the Loss Landscape of Neural Nets. NeurIPS.Li, H., Xu, Z., Taylor, G., & Goldstein, T. (2017).

Model Distillation

• Interpretable Mimic Learning

  Distilling knowledge from deep networks with applications to healthcare domain. Z. Che, S. Purushotham, R. Khemani, and Y. Liu. (2015).

• DarkSight

  Interpreting deep classifier by visual distillation of dark knowledge. K. Xu, D. H. Park, D. H. Yi, and C. Sutton. (2018).

• DeepVID

  DeepVID: Deep Visual Interpretation and Diagnosis for Image Classifiers via Knowledge Distillation. Wang, J., Gou, L., Zhang, W., Yang, H.T., & Shen, H. (2019). IEEE Transactions on Visualization and Computer Graphics, 25, 2168-2180.

Adverserial Attack

Adversarial examples: Attacks and defenses for deep learning. X. Yuan, P. He, Q. Zhu, and X. Li. (2017).

• AEVis

  Analyzing the Noise Robustness of Deep Neural Networks. M. Liu, S. Liu, H. Su, K. Cao, and J. Zhu. In Proceedings of the IEEE Conference on Visual Analytics Science and Technology, 2018.

  visualizes data-paths along the hidden layers in order to interpret the prediction process of adversarial examples.

  Explaining Vulnerabilities to Adversarial Machine Learning through Visual Analytics. Ma, Y., Xie, T., Li, J., & Maciejewski, R. (2019). IEEE Transactions on Visualization and Computer Graphics, 26, 1075-1085.

Feature Relevance/Importance Method

• Saliency

• Sensitivity

• Partial Dependence Plot

• Influence functions

  Understanding black-box predictions via influence functions, P. W. Koh, P. Liang, in: Proceedings of the 34th International Conference on Machine Learning. (2017)

• Interacticon based

  GoldenEye: A peek into the black box: exploring classifiers by randomization, A. Henelius, K. Puolamaki, H. Bostrom, L. Asker, P. Papapetrou, Data mining and knowledge discovery (2014)

  Interpreting classifiers through attribute interactions in datasets A. Henelius, K. Puolamaki, A. Ukkonen, (2017).arXiv:1707.07576.

• Others

  Iterative orthogonal feature projection for diagnosing bias in black-box models J. Adebayo, L. Kagal, (2016). arXiv:1611.04967.

Local interpretability

Global is Understanding of the whole logic of a model and follows the entire reasoning leading to all the different possible outcomes.

While local Explaining the reasons for a specific decision or single pre-diction

• LIME

  Why should i trust you?: Explaining the predictions of any classifier M. T. Ribeiro, S. Singh, and C. Guestrin,in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2016,

  Approximates a DNN’s predictions using sparse linear models where we can easily identify important features. Extracts image regions that are highly sensitive to the network output.

• LRP: Layer-Wise Relevance Propagation

  On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, and W. Samek, PLoS ONE, 2015.

• LOCO

  Distribution-free predictive inference for regression J. Lei, M. G’Sell, A. Rinaldo, R. J. Tibshirani, and L.Wasserman.

• Anchor

  Anchors: High-precision model-agnostic explanations M. T. Ribeiro, S. Singh, and C. Guestrin, in Proc. AAAI Conf. Artif. Intell., 2018.

Model Specific

Tree-based Model

• BoostVis

  Visual diagnosis of tree boosting methods. S. Liu, J. Xiao, J. Liu, X. Wang, J. Wu, and J. Zhu. IEEE Transactions on Visualization and Computer Graphics, 24(1):163–173, 2017.

• random forest

  iForest: Interpreting Random Forests via Visual Analytics Xun Zhao, Yanhong Wu, Dik Lun Lee, and Weiwei Cui. IEEE VIS 2018

  Summarize the decision paths in random forests.

  

CNN

1. Visualization

(1) Max Activation

Synthesize input pattern that can cause maximal activation of a neuron

• Saliency Maps(2013)

  Deep inside convolutional networks: visualising image classification models and saliency maps. Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. In arXiv:1312.6034, 2013.

  Saliency maps are usually rendered as a heatmap, where hotness corresponds to regions that have a big impact on the model’s final decision based on gradient.

  

• CAM: Class Activation Map(2016)

  The CAM highlights the class-specific discriminative regions.

  Learning Deep Features for Discriminative Localization. Zhou, B., Khosla, A., Lapedriza, À., Oliva, A., & Torralba, A.2016 IEEE (CVPR), 2921-2929.

  Note: https://medium.com/@ahmdtaha/learning-deep-features-for-discriminative-localization-aa73e32e39b2

  

  Grad-CAM: Why did you say that? R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, D. Batra,(2016).

• Filter Activation

  Convergent learning: Do different neural networks learn the same representations?, Y. Li, J. Yosinski, J. Clune, H. Lipson, J. E. Hopcroft, in: ICLR, 2016.

  computing the correlation between activations of different filters.

(2) Deconvolution(2010)

Finds the selective patterns from the input image that activate a specific neuron in the convolutional layers by projecting the lowdimension neurons'feature maps back to the image dimension

1. First propose Deconv

   Deconvolutional networks. M. D. Zeiler, D. Krishnan, G. W. Taylor, R. Fergus, in: CVPR, Vol. 10,2010, p. 7.

2. Use Deconv to visualize CNN

   Visualizing and understanding convolutional net-works. Matthew D. Zeiler and Rob Fer-gus. In ECCV, 2014.

(3) Inversion

Different from the above, which visualize the CNN from a single neuron’s activation, this methods is from Layer-level.

Reconstructs an input image based from a specific layer's feature maps, which reveals what image information is preserved in that layer

• Optimisation method

  Understanding deep image representations by inverting them. Aravindh Mahendran and Andrea Vedaldi. In CVPR, 2015.

  

  Visualizing deep convolutional neural networks using natural pre-images, A. Mahendran and A. Vedaldi, International Journal of Computer Vision, 120 (2016), 233–255.

• Up-conv net based

  Inverting visual representations with convolutional networks. Alexey Dosovitskiy and Thomas Brox. In CVPR, 2016.

  Train convolutional networks to reconstruct input images from different feature representations

  Object detectors emerge in deep scene cnns. Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. In ICRL, 2015.

  Reconstruct the minimal image representation that can be classified as the same category as the original image.

• Generative model

  Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Nguyen, A., Dosovitskiy, A., Yosinski, J., Brox, T., Clune, J.In: Advances in Neural Information Processing Systems, pp. 3387–3395 (2016)

  Plug & play generative networks: Conditional iterative generation of images in latent space. Anh Nguyen, Jeff Clune, Yoshua Ben-gio, Alexey Dosovitskiy, and Jason Yosinski. CVPR, 2017.

(4) Viusalization System: Understanding, Diagnosis, Refinement

• CNNVis

  Towards better analysis of deep convolutional neural networks, M. Liu, J. Shi, Z. Li, C. Li, J. Zhu, S. Liu, IEEE transactions on visualization and computer graphics 23 (1) (2016) 91–100.

  

  An Interactive Node-Link Visualization of Convolutional Neural Networks. Harley A.W. (2015) Advances in Visual Computing. ISVC

  showing not only what it has learned, but how it behaves given new user-provided input.

• Block

  Do convolutional neural networks learn class hierarchy? A. Bilal, A. Jourabloo, M. Ye, X. Liu, and L. Ren. IEEE transactions on visualization and computer graphics, 24(1):152–162, 2018.

  Including a class hierarchy and confusion matrix showing misclassified samples only, bands indicate the selected classes in both dimensions and a sample viewer

  

• DeepEyes

  A. DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks. Pezzotti, N., Höllt, T., van Gemert, J., Lelieveldt, B. P., Eisemann, E., & Vilanova, VAST 2017.

  Identification of stable layers, degenerated filters,patterns undetected, oversized layers, unnecessary layers or the need of additional layers

  

• Toolbox for visualization CNN

  Understanding Neural Networks Through Deep Visualization. Yosinski, J., Clune, J., Nguyen, A.M., Fuchs, T.J., & Lipson, H. (2015). ArXiv, abs/1506.06579.

  

  Visualizing the Hidden Activity of Artificial Neural Networks.Rauber, P.E., Fadel, S.G., Falcão, A.X., & Telea, A. (2017). IEEE Transactions on Visualization and Computer Graphics, 23, 101-110.

  Using dimensionality reduction for:

  1.visualizing the relationships between learned representations of observations

  2. visualizing the relationships between artificial neurons.

  

  Picasso: A Modular Framework for Visualizing the Learning Process of Neural Network Image Classifiers. Henderson, R. & Rothe, R., (2017). Journal of Open Research Software. 5(1), p.22.

  compute actual receptive field of filters.

2. Using explainable Model

• Decision Tree

  Interpreting CNNs via decision trees , Q. Zhang, Y. Yang, H. Ma, Y. N. Wu, in: IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 6261–6270.

3. Archtecture Modification

• All-conv net

  Striving for simplicity: the all convolutional net. ost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Ried-miller. ICLR workshop, 2015.

  Objext Detection:replace maxpooling layer with all conv-layers

• NIN

  Network In Network. Lin, M., Chen, Q., & Yan, S. (2013). CoRR, abs/1312.4400.

• LRCN

  Long-term recurrent convolutional networks for visual recognition and description. Donahue, J., Hendricks, L.A., Guadarrama, S., Rohrbach, M., Venugopalan, S., Saenko, K., & Darrell, T. (2014). IEEE Transactions on Pattern Analysis and Machine Intelligence, 39, 677-691.

RNN

• Visualization

  Visualizing and understanding recurrent networks A. Karpathy, J. Johnson, L. Fei-Fei, (2015).arXiv:1506.02078.

• Feature Relevence

  Explaining recurrent neural network predictions in sentiment analysis L. Arras, G. Montavon, K.-R. Muller, W. Samek, (2017). arXiv:1706.07206.

• RNNVis

  Understanding hidden memories of recurrent neural networks. Y. Ming, S. Cao, R. Zhang, Z. Li, Y. Chen, Y. Song, and H. Qu. In Visual Analytics Science and Technology (VAST), 2017 IEEE Conference on.IEEE, 2017.

• LISA

  LISA: Explaining Recurrent Neural Network Judgments via Layer-wIse Semantic Accumulation and Example to Pattern Transformation. Gupta, P., & Schütze, H. (2018). BlackboxNLP@EMNLP.

• LSTMVis

  Lstmvis: A tool for visual analysis of hidden state dynamics in recurrent neural networks. H. Strobelt, S. Gehrmann, H. Pfister, and A. M. Rush. IEEE transactions on visualization and computer graphics, 24(1):667–676,2018.

• Seq2Seq

  Seq2seq-Vis: A Visual Debugging Tool for Sequence-to-Sequence Models. Strobelt, H., Gehrmann, S., Behrisch, M., Perer, A., Pfister, H., & Rush, A.M. (2018). IEEE Transactions on Visualization and Computer Graphics, 25, 353-363.

• Attention

  Visualizing Attention in Transformer-Based Language models.Vig, J. (2019).

  Deep Features Analysis with Attention Networks. Xie, S., Chen, D., Zhang, R., & Xue, H. (2019). ArXiv, abs/1901.10042.

• Self-Attention

  SANVis: Visual Analytics for Understanding Self-Attention Networks. Park, C., Na, I., Jo, Y., Shin, S., Yoo, J., Kwon, B.C., Zhao, J., Noh, H., Lee, Y., & Choo, J. (2019). ArXiv, abs/1909.09595.

• Bert

  Visualizing and Measuring the Geometry of BERT. Coenen, A., Reif, E., Yuan, A., Kim, B., Pearce, A., Viégas, F.B., & Wattenberg, M. (2019). NeurlIPS, abs/1906.02715.

Generative Model

• GANViz

  GANViz: A Visual Analytics Approach to Understand the Adversarial Game. Wang, J., Gou, L., Yang, H.T., & Shen, H. (2018). IEEE Transactions on Visualization and Computer Graphics, 24, 1905-1917.

• DGTracker

  Analyzing the training processes of deep generative models. M. Liu, J. Shi, K. Cao, J. Zhu, and S. Liu. IEEE transactions on visualization and computer graphics, 24(1):77–87, 2018.

  

• GAN Lab

  GAN Lab: Understanding Complex Deep Generative Models using Interactive Visual Experimentation. Kahng, M., Thorat, N., Chau, D.H., Viégas, F.B., & Wattenberg, M. (2018). IEEE Transactions on Visualization and Computer Graphics, 25, 310-320.

Reinforcement Learning

• t-SNE

  Graying the black box: Understanding DQNs. Zahavy, T., Baram, N., and Mannor, S. International Conference on Machine Learning, pp. 1899–1908, 2016.

  using SAMDPs to analyze high-level policy behavior

• Saliency maps

  Visualizing and Understanding Atari Agents. Greydanus, S., Koul, A., Dodge, J., & Fern, A. (2017). ICML, abs/1711.00138.

  how inputs influence individual decisions using saliency maps

  

• Entropy

  Establishing appropriate trust via critical states. S. H. Huang, K. Bhatia, P. Abbeel, and A. D. Dragan.International Conference on Intelligent Robots (IROS), 2018

  finds critical states of an agent based on the entropy of the output of a policy.

• AM

  Finding and Visualizing Weaknesses of Deep Reinforcement Learning Agents. Rupprecht, C., Ibrahim, C., & Pal, C.J. (2019). ArXiv, abs/1904.01318.

  using activation maximization methods for visualization.

• DQNViz

  DQNViz: A Visual Analytics Approach to Understand Deep Q-Networks. Wang, J., Gou, L., Shen, H., & Yang, H.T. (2018). IEEE VAST, Transactions on Visualization and Computer Graphics(honorable mention), 25, 288-298.

  Extract useful action/reward patterns that help to interpret the model and control the training