Before a machine can perform intelli­gence tasks such as recognizing the details of an image, it must be trained. Training of modern-day artificial intelli­gence (AI) systems like Tesla’s autopilot costs several million dollars in electric power consumption and requires super­computer-like infra­structure. This surging AI appetite leaves an ever-widening gap between computer hardware and demand for AI. Photonic integrated circuits – optical chips – have emerged as a possible solution to deliver higher computing performance, as measured by the number of operations performed per second per watt used. However, though they’ve demons­trated improved core operations in machine intelligence used for data classi­fication, photonic chips have yet to improve the actual front-end learning and machine training process. 

Machine learning is a two-step procedure. First, data is used to train the system and then other data is used to test the performance of the AI system. Now, a team of researchers from the George Washington Univer­sity, Queens University, University of British Columbia and Princeton University set out to do just that. After one training step, the team observed an error and recon­figured the hardware for a second training cycle followed by additional training cycles until a suffi­cient AI perfor­mance was reached – e.g. the system is able to correctly label objects appearing in a movie. Thus far, photonic chips have only demons­trated an ability to classify and infer information from data. Now, researchers have made it possible to speed up the training step itself.

This added AI capability is part of a larger effort around photonic tensor cores and other electronic-photonic appli­cation-specific integrated circuits (ASIC) that leverage photonic chip manu­facturing for machine learning and AI appli­cations. “This novel hardware will speed up the training of machine learning systems and harness the best of what both photonics and electronic chips have to offer. It is a major leap forward for AI hardware acceleration. These are the kinds of advancements we need in the semi­conductor industry as underscored by the recently passed CHIPS Act”, Volker Sorger from the George Washington University said. “The training of AI systems costs a signi­ficant amount of energy and carbon footprint. For example, a single AI trans­former takes about five times as much CO₂ in elec­tricity as a gasoline car spends in its lifetime. Our training on photonic chips will help to reduce this overhead”, Bhavin Shastri from Queens University added. (Source: GWU)

Reference: M. J. Filipovich et al.: Silicon photonic architecture for training deep neural networks with direct feedback alignment, Optica 9, 1323 (2022); DOI: 10.1364/OPTICA.475493

Link: Devices & Intelligent Systems Laboratory, Dept. of Electrical and Computer Engineering, George Washington University, Washington DC, USA

Further Reading: A. Gonzalez & D. Costa: Programmable chips for 5G applications, PhotonicsViews 19, 38 (2022); DOI: 10.1002/phvs.202200002