Explainable machine learning

elective monographs

Deep neural networks 1000-317bDNN
Statistical machine learning 1000-317bSML

The goal of the course is to learn about concepts, methods and techniques for explaining complex machine learning models. Predictive models are becoming increasingly complex, tree ensembles, deep neural networks are models with thousands of parameters. For models with such dimensionality, it is easy to lose track of what the model has learned.

During this course we will discuss tools for analysing the structure of a model treated as a black box, and analysing the predictions from the model. This will allow us to increase confidence in the model, improve model performance, and be able to extract useful knowledge from the model. We will learn about the most popular explanatory methods, discuss their strengths and weaknesses so that the class participant has the necessary competences to further explore the literature in this area.

- Introduction to explainable artificial intelligence, interpretable machine learning and fairness

- Methods for conditional analysis of predictive models: Break-Down method, Break-Down with interactions, SHAP, ASV

- Methods for model analysis by perturbation: LIME method, LORE

- Methods for contenst model analysis and model sensitivity testing: Ceteris Paribus, Partial Dependence, Accumulated Local Methods

- Method for assessing the importance of variables: Variable Importance by Pertmutations, Model Class Reliance

- Fairness and Biases

- Explanations specific to neural networks