EAGER:Scalable Photonic AI Accelerators Based on Photoelectric Multiplication

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

One of the deepest questions in science is how biological cognition works. Traditionally the purview of neuroscience and psychology, in recent years computer science have shed light on it through the field of 'deep learning'. Deep learning uses computer algorithms called neural networks to perform various tasks-e.g. face recognition, medical diagnosis, automobile driving--that have long been considered difficult for computers, and is transforming many industries including logistics, manufacturing, healthcare, and finance. However, neural networks are very costly to run even on modern computers. To unlock deep learning's full potential, this research will investigate a new concept: Optical Neural Networks. By running neural networks on dedicated optical hardware, there is a potential to increase speed and reduce energy consumption by at least 1000x. This program will study the feasibility of this concept to pave the way for more extensive technology development in the future. If realized, Optical Neural Networks will allow researchers to develop significantly larger, more complex deep learning models that may open up entirely new deep learning applications that are beyond the capabilities of present-day computers. Artificial intelligence (AI) based on deep neural networks (DNNs) has revolutionized a wide range of fields, but at a cost: DNNs are very compute- and power-intensive. Driving the AI revolution has been an exponential growth in the available compute performance, which has enabled the application of DNNs to increasingly complex tasks. However, as Moore's Law runs out of steam, this trend cannot continue for long; therefore, the development of alternative platforms for AI hardware has become especially urgent. This EAGER will study a class of optical neural networks (ONNs) that harness the unique advantages of photonics and promise orders-of-magnitude throughput- and energy-consumption improvements over conventional electronics. Three key tasks are: (i) a system-level architecture study to predict the ONN's performance gains on realistic workloads, (ii) a hardware analysis and feasibility study, and (iii) an investigation into the fundamental limits of ONNs. Research activities include modeling, numerical analysis, and benchmarking. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.