Introduction to eXplainable Machine Learning (XAI)

Date:

June 19 to 23. MORNING: 9 to 12h

Instructor

Przemyslaw Biecek’s goal is to support humans’ effectiveness through safe, ethical, effective and automated predictions. He implement this by developing processes, methods, tools, and software for responsible machine learning.

He defended his PhD in mathematical statistics in 2007 and gained the full professor title in computer science in 2023. During this time he worked in various positions. As academic researcher at the Warsaw University of Technology and the University of Warsaw. As AI expert at the OECD and GPAI.AI. As ML specialist at Samsung, IBM, Netezza and Disney. As entrepreneur, he founded the MI2.AI RedTeam which offers services on training and auditing predictive models from the perspective of transparency, robustness and fairness.

He has always been fascinated by data visualisation. He now uses this interest to work on the visualisation of predictive models. This is the topic of his latest book “Explanatory Model Analysis” https://ema.drwhy.ai/.

In free time, he writes stories and comics in the Beta and Bit series introducing Data Literacy to high school students https://www.mi2.ai/beta-bit.html.

Language

English

Description

Introduction to eXplainable Machine Learning (XAI)

A course consisting of lectures and hands-on workshops and a short project on the topic of explainable machine learning.
During the course we will discuss key techniques for visual exploration of predictive models.

Course goals

Participants are familiar with the strengths and weaknesses of key predictive model exploration techniques (SHAP, LIME, Partial Dependence, ALE).
Participants are able to apply them in practice to analyze a selected predictive problem (e.g., classification).

Course cover material from books Explanatory Model Analysis and The Hitchhiker’s Guide to Responsible Machine Learning

Course contents
Day 1:

Lecture: Introduction to eXplainable Machine Learning, why it is important.

Lab: Development of ML models that will be later used for model analysis

Day 2:

Lecture: Introduction to local methods LIME and Break-Down / SHAP

Lab: Analysis of selected models with methods for local attribution

Day 3:

Lecture: Introduction to Ceteris Paribus / Partial Dependence / ALE methods

Lab: Analysis of selected models with methods for profile analysis

Day 4:

Lecture: Introduction to Variable Importance and Fairness

Lab: Analysis of selected models with methods for variable importance

Day 5:

Presentation of projects

Prerequisites

Basic knowledge of R (or Python)
Basic experience in predictive modelling (logistic regression or random forest classification will be a big plus)

Targeted at

Students in mathematics, computer science or other fields with computational methods, and practitioners who want to use predictive models, are interested in visualizing these models and better understanding the strengths and weaknesses of these models.

Evaluation

The methods presented in the first four days will be used to prepare a project (short report with a presentation).
The results of the projects will be presented on the last day.
The projects will be the basis for evaluation.

Software requirements

R in version 4.2 or higher

RStudio or other editor

Packages: DALEX, mlr3, randomForest, ggplot2, partykit