High-Precision Analog Photonic Accelerators for Artificial Neural Networks

Abstract

Abstract:We propose a photonic tensor accelerator based on coherent mixing and multi-dimensional accumulation for artificial neural networks. Our preliminary work has shown that the photonic tensor accelerator can potentially provide unprecedented scale, speed (single-clock cycle) and energy efficiency for matrix-vector, matrix-matrix, and tensor-tensor multiplications by exploiting the wavelength, mode and spatial dimensions of light. In this work, we propose to enhance the accuracy of the photonic tensor accelerator by reducingthe crosstalk among multiplication products using two-dimensional (wavelength and mode) orthogonality during accumulation. In addition, we propose to implement the photonic tensor accelerator on silicon photonic (SiPh) integrated circuits, rather than in free space, by taking advantage existing single-modefoundry services as well as developing SiPh components for both wavelength and mode multiplexing. The scale and efficiency offered by the photonic tensor accelerator are not available today and are not expected to be available any time soon in electronics. Therefore, this research may well become the first step forphotonics to play major roles in artificial neural networks.

Document Details

Document Type
DoD Grant Award
Publication Date
Jun 17, 2020
Source ID
N000142012441

Entities

People

  • Guifang Li

Organizations

  • Office of Naval Research
  • United States Navy
  • University of Central Florida Board of Trustees

Tags

Readers

  • Integrated Circuit Design and Technology.
  • Linear Algebra
  • Pulsed Power and Plasma Physics.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Microelectronics
  • Space
  • Space - Hall-Effect Thruster