Tutorials for eXplainable Artificial Intelligence (XAI) methods

Complex supervised Machine Learning (ML) and Deep Learning (DL) models are usually trained to achieve high accuracy, but are often considered to be “Black Boxes” because of a lack of understanding of the underlying reasons for decisions made by the model. To deploy such models to the real world, i.e. to move from research to application, it is indispensable to not only make accurate predictions, but also to understand the logic behind those predictions, to ensure that the models produce valuable results and are safe to use [1]. The application of interpretability methods allows us to identify potential biases, ensure fairness and prevent genre, social, and race discrimination. Moreover, in many fields, especially in the biomedical domain, explainability can be the key to acceptance and trust.

The XAI-Tutorials repository provides a collection of self-explanatory tutorials for different model-agnostic and model-specific XAI methods. Each tutorial comes in a Jupyter Notebook with practical exercises. In addition, we provide useful background information on those methods in this documentation.

EXPLAINABLE AI

• Terminology
• Why is explainability important?
• Taxonomy of XAI methods

MODEL-AGNOSTIC XAI METHODS

• Introduction to Permutation Feature Importance
• Introduction to Local Interpretable Model-Agnostic Explanations (LIME)
• Introduction to SHapley Additive exPlanations (SHAP)

MODEL-SPECIFIC XAI METHODS

• Introduction to Forest-Guided Clustering
• Introduction to Grad-CAM
• Introduction to Attention Maps

MACHINE LEARNING BASICS

• Introduction to Random Forest Models
• Introduction to Convolutional Neural Networks and Feature Visualization
• Introduction to Transformers

Contributions

Comments and input are very welcome! Please, if you have a suggestion or you think something should be changed, open an issue or submit a pull request.

References

[1] R. Roscher, B. Bohn, M. F. Duarte and J. Garcke, “Explainable Machine Learning for Scientific Insights and Discoveries,” IEEE Access, Vol. 8, pp. 42200-42216 (2020)