Mathematical mysteries of deep neural networks
Apparaît dans la collection : AMS-EMS-SMF 2022
Deep neural networks have spectacular applications to classification, regression or generation of images, sounds, languages and most physical data. They are responsible of the renewal of artificial intelligence. Yet, we mostly do not understand the underlying mathematics, allowing them to avoid the curse of dimensionality. By taking examples and inspiration from statistical physics, this talk willemphasize the role of concentration phenomena, multiscale interactions and invariants. Applications will be shown in physics and for image classification problems.
