Metrics and a cut-off phenomenon in deep learning (Anders Karlsson, UNIGE)

29.11.2023 10:30

Today’s AI is based on neural networks and deep learning. These concepts are easily explained to a mathematician. Why and exactly how it works are however much more unclear, and the problems and risks associated with this are widely discussed. Products of random transformations occur in practice in several ways in deep learning (initialization, stochastic gradient descent, drop-out, etc). There is a noncommutative ergodic theorem that governs such products asymptotically. For less large neural networks we also found empirical evidence of a cut-off phenomenon for random initialization. Ongoing work with B. Avelin (Uppsala) and B. Dherin (Google).

Lieu

Bâtiment: Conseil Général 7-9

Room 1-07, Séminaire d'analyse numérique

entrée libre