Abstract:Deep neural networks have achieved remarkable breakthroughs by leveraging multiple layers of data processing to extract hidden representations, albeit at the cost of large electronic computing power. To enhance energy efficiency and speed, the optical implementation of neural networks aims to harness the advantages of optical bandwidth and the energy efficiency of optical interconnections. In the absence of low-power optical nonlinearities, the challenge in the implementation of multilayer optical networks lies in realizing multiple optical layers without resorting to electronic components. In this study, we present a novel framework that uses multiple scattering that is capable of synthesizing programmable linear and nonlinear transformations concurrently at low optical power by leveraging the nonlinear relationship between the scattering potential, represented by data, and the scattered field. Theoretical and experimental investigations show that repeating the data by multiple scattering enables non-linear optical computing at low power continuous wave light. Moreover, we empirically found that scaling of this optical framework follows the power law as in state-of-the-art deep digital networks.

What problem does this paper attempt to address?

This paper aims to solve the implementation problems of linear and nonlinear calculations in optical neural networks (ONNs), especially under low - power consumption conditions. Specifically, the main challenge faced by researchers is how to complete nonlinear calculations without relying on electronic components when implementing multi - layer networks in the optical domain. Although a variety of methods have demonstrated efficient optical linear - calculation hardware, effectively combining these capabilities with nonlinear calculations remains a significant obstacle. In addition, existing optical - calculation schemes are often limited by non - programmable optoelectronic nonlinearity or high energy consumption. To solve these problems, the author proposes a framework named "Nonlinear Processing with Only Linear Optics" (nPOLO). nPOLO utilizes low - power continuous - wave lasers and diffraction layers to simultaneously achieve linear and nonlinear operations in the optical domain through multiple scatterings. The core of this framework lies in using multiple data planes and forming a multi - reflection single - pass cavity through the configuration of a spatial light modulator (SLM) and mirrors, with each plane modulating the phase of light as a reflecting surface. In this way, nPOLO can generate high - order nonlinear transformations of data, thereby achieving complex optical neural network calculations while maintaining low power consumption. This research not only verifies the effectiveness of the nPOLO framework in synthesizing learnable linear and nonlinear data transformations but also demonstrates its classification performance on different datasets through experiments and simulations. The results show that increasing the number of layers and the number of data repetitions can increase the order of nonlinearity, thereby improving the accuracy of the model and its robustness to experimental imperfections. In addition, the study also finds that the performance of nPOLO improves with the increase in the number of parameters, exhibiting a power - law scaling characteristic similar to that of the state - of - the - art digital deep networks. These findings provide new ideas and technical paths for the implementation of all - optical neural networks.

Nonlinear Processing with Linear Optics

Nonlinear optical encoding enabled by recurrent linear scattering

Low-power scalable multilayer optoelectronic neural networks enabled with incoherent light

Deep Learning with Passive Optical Nonlinear Mapping

Fully Non-Linear Neuromorphic Computing with Linear Wave Scattering

Nonlinear encoding in diffractive information processing using linear optical materials

Large-scale photonic computing with nonlinear disordered media

Nonlinear Optics for High-Speed Digital Information Processing

All-Optical Computing Based on Convolutional Neural Networks

InP Photonic Integrated Multi-Layer Neural Networks: Architecture and Performance Analysis

Advances and Marine Applications of Optical Neural Network

Lensless Opto-Electronic Neural Network with Quantum Dot Nonlinear Activation

An Optical Neural Network Architecture based on Highly Parallelized WDM-Multiplier-Accumulator

Training neural networks with end-to-end optical backpropagation

Lensless opto-electronic neural network with quantum dots nonlinear activation

Genetically programmable optical random neural networks

On the implementation of nonlinearities in optical neural networks: opinion

At the intersection of optics and deep learning: statistical inference, computing, and inverse design

Optical neural networks: progress and challenges

Nonlinear unitary circuits for photonic neural networks

Forward-Forward Training of an Optical Neural Network