Demonstrating completeness in optical neural computing

TL;DR

This paper demonstrates a complete optical neural network using silicon photonics for fast and energy-efficient computing.

Contribution

The novel contribution is an integrated optical neural network with on-chip nonlinear activations for end-to-end inference.

Findings

The optical neural network uses partially coherent light for high-speed inference.

The system integrates convolutional and fully connected layers with optoelectronic nonlinear activations.

It provides a scalable platform for evaluating optical neural processing beyond traditional electronics.

Abstract

A silicon photonic deep optical neural network integrating convolutional and fully connected layers with on-chip optoelectronic nonlinear activations operates with partially coherent light to achieve high-speed, energy-efficient, end-to-end inference. This demonstration establishes a functional and scalable platform for evaluating complete optical neural processing, representing another step toward specialised, ultrafast photonic architectures beyond electronics.

Figures