Oberseminar "Mathematik des Maschinellen Lernens und Angewandte Analysis" - Atell Yehor Krasnopolsky

Sensitivity maps are gradient-based explainability tools used in machine learning to visualise the parts of the input most influential to a model’s decision-making. Training a model to be robust against slight perturbations of the input (adversarial training) has an effect on its sensitivity maps: they become sparser. The goal of my thesis was to mathematically answer the question: why is this the case? In this talk I will share the results.