Neuromorphic Computing with Silicon Photonics

Abstract

Modern computing has been marked by a fascinating exponential trajectory since the invention of the transistor. Global networking and the Internet have also experienced an exponential growth thanks to optical fiber networking. Today, due to advances in integrated photonics, optical systems can overcome barriers in high speed computing beyond the capabilities of semiconductor electronics. Neuromorphic photonic computing draws from an improbable combination of three well-established research fields: silicon photonics, neuromorphic computing, and machine learning. This multidisciplinary field seeks to engineer efficient hardware that can emulate brain-inspired neural networks, which power most artificial intelligence applications today. By combining photonic devices and neuro-inspired architectures, this hardware can perform AI processing computations at very high speeds.There is a fundamental reason for drawing inspiration from neuroscience for optical computing. The physics of photonic interactions is much more similar to the distributed nature of neural networks than the serial nature of digital processors. Machine learning applications, for example, benefit from having thousands of interconnections between neurons, which can fit onto a single optical waveguide. The density of electrical interconnects between digital circuits, on the contrary, is severely limited at high frequencies. This presents a unique opportunity to engineer a neuromorphic photonic processor that can advance the state-of-the-art signal processing bandwidth by orders of magnitude. The primary result of this thesis is a demonstration of a fully reconfigurable photonic neural network integrated on chip. While neural network functionalities have been individually shown before, the system integration on a single chip that we show here is a formidable challenge that requires many systemic tradeoffs with many potential points of failure. We first provide an overview of the demonstrations of individual photonic devices in multiple fabrication platforms. We then report the results of neural network experiments, which can be generalized to real-world applications in real-time computing, signal processing, optical communications, and predictive control. To foster wide adoption beyond the photonics community, we provide a blueprint for programming the chip with the help of digital electronics. Finally, we show that photonic neural networks are scalable without runaway noise accumulation, provided that efficient modulators and light sources are used.