Advanced ML: Interpretable and Explainable AI and Uncertainty Estimations

by Maciej Mikolaj Glowacki (CERN)

Monday 24 Aug 2026, 09:00 → 16:00 Europe/Zurich

593/R-010 - Salle 11 (CERN)

Abstract

This one-day course covers both sides of trustworthy AI in physics: understanding what a model has learned, and knowing how far to trust its predictions. The first lecture shows ways to interpret neural networks: feature attribution (SHAP, Grad-CAM, attention), loss-landscape analysis, and opening the “black box” through latent-space geometry and sparse autoencoders. The second lecture shows how to quantify and validate model uncertainty, using deep ensembles, Bayesian neural networks and conformal prediction, with checks of coverage, calibration and bias. It includes an ATLAS calorimeter case study. A hands-on practice session puts both topics into use.