Deep physical neural networks: training physical systems like neural networks

Department of Applied Physics
Solid State & Optics Seminar Series

Sponsored by "The Flint Fund Series on Quantum Devices and Nanostructures"

Dr. Logan Wright
Wednesday, March 23, 2022
1:00 PM
Via Zoom
Meeting ID: 947 2015 4098
Password: 604783
https://yale.zoom.us/j/94720154098?pwd=N- 1hvZUQvWng2a2lZdndVVldDT1FzQT09

Deep physical neural networks: training physical systems like neural networks

Deep learning has proven to be a remarkably versatile and scalable technique for learning algorithms to process and interact with noisy, high-dimensional real-world data and systems. In deep learning, the backpropagation algorithm is used to adjust the parameters of a multi-layer (deep) neural network so that the network "learns" to perform desired mathematical functions. Here, I will discuss my work to adapt this procedure to train networks of controllable physical systems – physical neural networks (PNNs) - which directly learn physical functions, such as performing machine learning inference calculations. I will present proof-of-concept PNNs we have con- structed to perform image and audio classification, based on ultrafast nonlinear photonics, bulk analog electron- ics, and mechanics. Because PNNs learn physical transformations directly, without relying on rigid mathemat- ical isomorphisms, they may harness noisy, analog physical processes for computation more opportunistically than traditional approaches. For example, PNNs based on nonlinear optical waves or microwave oscillators offer routes to performing machine learning calculations millions of times faster and more energy-efficiently than conventional hardware. More broadly, PNNs form the basis for a learning-based approach to the design and programming of complex physical devices, such as optical sensors that perform ultrafast signal processing to drastically reduce the time, dose, and cost of biomedical diagnostics.

Hosted By: Professor Hui Cao