Princeton University

School of Engineering & Applied Science

Silicon Photonic Neural Networks

Speaker: 
Alexander Tait
Advisor: 
Prof. Prucnal
Location: 
Engineering Quadrangle J401
Date/Time: 
Monday, January 22, 2018 - 10:30am to 12:00pm

Abstract
Today’s computing demands approach the limits of microelectronics. In radio signal processing, real-time control, and some mathematical programming areas, it will become necessary to use new device physics and new computer architectures. Photonic devices are extremely efficient for high-bandwidth processing and have already found simple analog signal processing niches. In order to solve more complex information processing problems, they will have to adopt a more general processing model: neural networks. Neural network models have come to dominate modern machine learning algorithms, and specialized electronic hardware has been developed to implement them more efficiently.
 
Neuromorphic photonics unites the innovations of two growing fields, neural networks and silicon photonics, to present a potential computing platform free from the specific limitations of conventional microelectronics. Neuromorphic photonics is unconventional in that it is continuous-valued, decentralized, and photonic (i.e. optical). On a silicon chip, neurons (instead of gates) are densely interconnected using light (instead of electricity). An advantage of being neuromorphic is that existing theory, such as machine learning, can be applied; an advantage of being compatible with silicon photonics is that existing photonic foundry platforms can be used.
 
We introduce the microring weight bank as the novel device that configures connection strengths between photonic neurons on a silicon chip, and we demonstrate small neuromorphic photonic systems. This research takes neuromorphic photonics, for the first time, from isolated photonic neuron devices to networks of photonic neurons. It further describes the design principles of these silicon photonic neural networks and provides example benchmark tasks for measuring the platform’s processing capabilities. In doing so, it builds the foundation for the continued exploration of the bright future of neuromorphic photonics.