Event

Conventional digital computing technology based on complementary metal–oxide–semiconductor (CMOS) devices has accumulated a performance increase in excess of 1,000,000x versus the first CMOS digital processors. However, the demand for more computing power (and simultaneously increased energy efficiency) – especially for neural networks and for processing a deluge of sensor data – has increased far more rapidly than what even the already-brisk pace of advances in modern processors and special-purpose accelerator chips have been able to deliver. How can we meet this demand over the next decade and beyond?

We study how, by relaxing or breaking some of the tenets of modern computing – such as digital signals, deterministic behavior, computational universality, and a hierarchy of abstractions from digital logic through to software – and reimagining how computers are built at the fundamental level of physical dynamics, we may be able to make orders-of-magnitude leaps in speed or efficiency in specialized co-processors.

In this talk I will briefly introduce the general research program of physics-based computing and then focus on a concrete example, physical neural networks [1]. I will describe a method my group has developed to train any complex physical system to perform as a neural network for machine-learning tasks, and then give an application of this method to a new kind of on-chip photonic neural processor we have developed: one whose refractive index as a function of space we can reprogram with light [2]. I will also discuss the potential of physical neural networks for smart sensors that pre-process acoustic, microwave or optical signals in their native domain before digitization [3].

[1] L.G. Wright*, T. Onodera* et al. Nature 601, 549-555 (2022)

[2] T. Onodera*, M. Stein* et al. arXiv:2402.17750 (2024)

[3] T. Wang*, M. Sohoni* et al. Nature Photonics 17, 408 (2023)