Using Photonics to Accelerate AI and Machine Learning Tasks

[Labs_Editorial]

In recent years research labs have done extensive work trying to get optical neural networks to accelerate the ever-growing number of computational tasks required by artificial intelligence and machine learning applications. It’s been an uphill battle.

Current technology solutions are held back by three issues: 1) a lack of high-speed and power-efficient devices that enable the large matrix multiplications that drive the computations; 2) a lack of on-chip optical memory; and 3) a proper material platform that can seamlessly integrated all the necessary optical components on a monolithic chip.

Now Hewlett Packard Labs has come up with a new way to tackle these issues. In a recent paper published in Nature Communications Engineering, a team in the Large-Scale Integrated Photonics (LSIP) Lab demonstrates how photonics can speed up the training and inference processes of AI applications.    

Stanley Cheung“These advancements pave the way for the development of future optical in-memory computing architectures and programmable photonics,” said Hewlett Packard Labs Principal Research Scientist Stanley Cheung, the lead author of the paper. “We’ve been able to show two important things: that it’s possible to perform high-speed and energy efficient training of AI models with our III-V/Si phase shifters and that our device can act as a non-volatile photonic memory element allowing multiplication with near zero static power consumption.”

The paper, Energy Efficient Photonic Memory Based on Electrically Programmable Embedded III-V/Si Memristors: Switches and Filters, was published March 18. Cheung co-wrote the paper with seven Labs researchers and collaborators: Bassem Tossoun, Yuan Yuan, Yingtao Hu, Wayne V. Sorin, Geza Kurczveil, Di Liang and Raymond G. Beausoleil.

Researchers in Hewlett Packard Labs’ LSIP Lab have published two other recent papers in Nature Communications advancing aspects of optical technology. One paper focused on innovations in high-bandwidth optical communications. Another paper lays out a complimentary approach to speeding up matrix multipliers.

In the latest paper, the team uses a device configuration called a Mach-Zehnder Interferometer that allows it to be used in a mesh architecture which can possibly scale toward high-intensity applications.

“We believe these photonic memristors can contribute to the kind of future energy efficient, non-volatile, large-scale integrated photonics uses that can be so important to society,” Cheung said. “These include neuromorphic/brain inspired optical networks, optical switching fabrics for tele/data-communications, optical phase arrays, quantum networks, and future optical accelerator architectures.”