Spatially Parallel All-optical Neural Networks

TL;DR

This work tackles the limits of all-optical neural networks that rely on serial layer arrangements by introducing a spatially parallel architecture. By splitting inputs into multiple parallel optical paths and coherently recombining their outputs, SP-AONNs generate nonlinearity through interference without active nonlinear components, enabling scalable, low-noise computation. Hardware demonstrations using a modular 4F system show that adding parallel sub-networks improves accuracy and robustness across several image benchmarks and withstands phase noise much better than serial counterparts. The results establish spatial parallelism as a practical route to high-capacity, energy-efficient optical neural computing with potential extensions to on-chip photonics and partially coherent light sources.

Abstract

All-optical neural networks (AONNs) have emerged as a promising paradigm for ultrafast and energy-efficient computation. These networks typically consist of multiple serially connected layers between input and output layers--a configuration we term spatially series AONNs, with deep neural networks (DNNs) being the most prominent examples. However, such series architectures suffer from progressive signal degradation during information propagation and critically require additional nonlinearity designs to model complex relationships effectively. Here we propose a spatially parallel architecture for all-optical neural networks (SP-AONNs). Unlike series architecture that sequentially processes information through consecutively connected optical layers, SP-AONNs divide the input signal into identical copies fed simultaneously into separate optical layers. Through coherent interference between these parallel linear sub-networks, SP-AONNs inherently enable nonlinear computation without relying on active nonlinear components or iterative updates. We implemented a modular 4F optical system for SP-AONNs and evaluated its performance across multiple image classification benchmarks. Experimental results demonstrate that increasing the number of parallel sub-networks consistently enhances accuracy, improves noise robustness, and expands model expressivity. Our findings highlight spatial parallelism as a practical and scalable strategy for advancing the capabilities of optical neural computing.

Figures (4)

• Figure 1: (a) Schematic of the serial all-optical Fourier neural network architecture. (b) Schematic of the parallel all-optical Fourier neural network architecture.(c) Schematic of the coherent-induced nonlinear computational process. (d) Neuron counts in parallel and serial architectures with varying layer numbers.
• Figure 2: (a-c) Simulated accuracy of AONNs with parallel and serial architectures at different layers. (d-f) Experimental accuracy of AONNs with parallel and serial architectures at different layers. (g-i) Experimental classification accuracy when blocking varying numbers of sub-networks in deployed N=4 SP-AONNs
• Figure 3: (a) Experimental schematic of the SP-AONNs. DMD: digital micromirror device loaded with input images; SLM1: spatial light modulator loaded with Dammann grating map; SLM2: spatial light modulator loaded with a trained phase map; FL: Fourier lens employed as the device for the Fourier transform of the light field. CCD: charge-coupled device employed to capture the output light field. (b) Training schematic of SP-AONNs. (c1-e2) Simulated and experimental output light fields of SP-AONNs, with insets magnifying classification regions to reveal interference-generated lattice patterns.
• Figure 4: Classification accuracy of N=4 parallel and serial AONNs under varying phase noise levels.