Prof. Paolo Missier – University of Birmingham, UK
Part I of this seminar series runs over 10 hours and is focused on Explainable AI (XAI), covering the following topics:
– The need for explanations in Machine Learning and AI
– Classical global and local methods : LIME and Shapley (Shap values)
– Influence functions: a robust statistical approach that has recently (circqa 2017) been revisited to associate a relative importance to training examples for explaining a specific model inference. One key advantage of using influence analysis over other methods, is its ability to bypass the “black box” barrier that is typical of complex nonlinear models (eg deep neural networks).
The programme includes:
– About 4 hours of seminar-style lectures, using key papers and a recent survey to dive into specific approaches to using influence functions
– Guided study to relevant literature, where students are required to select recent work from a portfolio of recommended papers, and to present their insights to the class
– Practical work using python libraries that implement influence-based methods to provide explanations, in combination with simple ML models.
Class schedule (All classes will take place in the Math building of the FIM Department, via Campi 213/B Modena)
Mon 12/5 h.12:00-14:00 – classroom M2.3
Tue 13/5 h.11:00-13:00 – 14:00-16:00 – sala Riunioni piano uffici
Thu 15/5 h. 11:00-13:00 – 14:00-16:00 – sala Riunioni piano uffici
The flyer for the event is available at this link.