A new kind of optical computing hardware for Artificial Intelligence
Building an all-optical computer is a long-standing goal in modern research, as the speed of computation could be increased by several orders of magnitude. Most of the attempts of achieving this goal have failed so far; however, optics remain a very promising platform to achieve some types of parallel computation. For example, the glasses you are wearing right now to read this article are continuously performing a Fourier transform, without any energy consumption. Researchers from the group I have joined this year (Marin Soljacic, Photonics and Modern Electromagnetics Group, in collaboration with Dirk Englund at MIT) have demonstrated a fully-optical platform to perform basic machine learning tasks (here : vowel recognition).
Photonics and Modern Electromagnetics Group
As the paper was published in Nature Photonics , it is a very promising proof-of-concept as, for instance, huge data centers require efficient architectures to process gazillions of data a day. An optical architecture leverages on the speed and the potential of massive processing parallelization, with a very low-power hardware.
I am currently working on a follow-up of this project to use the power of optics to solve “hard” computational problems.
Some other news articles on this article and on the startup:
Charles Roques Carme – Carnot fellow 2016