Jacobians of neural networks and Free Probability

Reda Chhaibi (IMT - Université Paul Sabatier)

30 septembre 2021, 11h00–12h15


Salle A4

MAD-Stat. Seminar


Gradient descent during the learning process of a neural network can be subject to many instabilities. The spectral density of the Jacobian is a key component for analyzing robustness. Following the works of Pennington et al., such Jacobians are modelled using free multiplicative convolutions from Free Probability Theory. In a substantial first part of the talk, I will present the problem from Machine Learning and introduce the necessary tools from Free Probability. It will then be clear that a good computational method is missing. In the second part of the talk, we present our solution: a reliable and very fast method for computing the spectral densities associated to a neural network. Beyond machine learning, the method is certainly of independent interest in Free Probability Theory and high dimensional statistics. This is joint work with Tariq DAOUDA and Ezéchiel KAHN.

Voir aussi