Buckley, S.
, McCaughan, A.
and Oripov, B.
(2024),
Photonic Online Learning, Journal of Physics: Photonics, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=957961
(Accessed January 17, 2025)
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
Secure .gov websites use HTTPS
A lock (
) or https:// means you’ve safely connected to the .gov website. Share sensitive information only on official, secure websites.
Photonic Online Learning
Published
Author(s)
, ,
Abstract
Training in machine learning necessarily involves more operations than inference only, with higher precision, more memory, and added computational complexity. In hardware, many implementations side-step this issue by designing "inference-only" hardware that is trained separately in a one-time simulation. The resulting weights and biases from the training simulation are then transferred to hardware. This is called "offline" or "in-silico" training. While this approach is well-suited to digital systems, in systems with analog components there is often significant degradation in performance accuracy between the simulation and the hardware implementation due to noise, device-to-device variations and drift. In addition, once trained, a new simulation is required if the application or hardware parameters change over time. A promising alternative approach is "online learning". We define online learning as any training process that involves making measurements on the physical system itself during training. Online learning techniques have enabled experimental demonstrations of photonic networks that can solve large-scale problems.Citation
Journal of Physics: Photonics
Pub Type
Journals
Download Paper
Citation
Issues
If you have any questions about this publication or are having problems accessing it, please contact reflib@nist.gov.
Created January 9, 2024, Updated January 3, 2025