Coffee break will be served before the Seminar at 10:30 on the Leonardo Building terrace.

- - -

Florian Marquardt
(the Max Planck Institute for the Science of Light and Friedrich-Alexander-Universität Erlangen-Nürnberg)

The ongoing revolution in artificial intelligence requires ever-increasing resources for training and deploying artificial neural networks. This exponentially accelerating trend has been recognized as unsustainable, and a community of scientists and engineers are exploring energy-efficient alternatives to the current paradigm of digital artificial neural networks. By designing devices that operate much closer to the microscopic laws of physics but offer similar functionalities, one can achieve significant gains in efficiency and speed. In this talk, I will give an overview of the different approaches. Above all, I will concentrate on the topic of physics-based learning, where the idea is to extract the correct updates to trainable parameters in an efficient manner, based on intelligently chosen experimental procedures. I will illustrate the opportunities in this domain by a few of our recent ideas. These include learning with the help of time-reversal operations, nonlinear processing based on linear optical scattering, and a new training technique that will help to turn any analog quantum simulator platform into a trainable neuromorphic device.
 

Zoom registration link:
https://zoom.us/meeting/register/RKsl8-NEQHuLifpYFYgE5A