TRON: Trainable, architecture-reconfigurable random optical neural networks

TL;DR

TRON is a scalable, trainable optical neural network that uses a multi-scattering medium and a DMD for high-dimensional matrix multiplication, enabling in-situ optimization and neural architecture search on optics.

Contribution

It introduces a novel optoelectronic deep optical neural network with in-situ optimization and NAS, advancing scalable reconfigurable optical computing for machine learning.

Findings

In-situ NAS is crucial for architecture adaptation to tasks and hardware.

TRON demonstrates effective training and optimization of optical neural networks.

The approach paves the way for large-scale optical processors in data-intensive applications.

Abstract

Deep learning has triggered explosive growth in the demand for specialized hardware processors, thus motivating the development of scalable and reconfigurable computing substrates. Optical processors offer a fundamentally different computing paradigm, combining massive parallelism and ultrahigh bandwidth with the potential for substantial energy savings. However, progress has been constrained by the absence of scalable and reconfigurable architectures that can implement a broad class of network architectures. Here, we introduce TRON, a scalable and trainable optoelectronic deep optical neural network that exploits a multi-scattering medium and a DMD as a learnable, high-dimensional dense optical matrix multiplier, processing with fixed and tunable optical operations. We perform in-situ optimization of the optical parameters involved in the scattering process, together with automated neural architecture search (NAS) and optimization directly on optics. The experimental results demonstrate that in-situ NAS is essential to discover architectures that adapt to both the task and hardware constraints, establishing a viable path towards large-scale optical processors for next-generation machine learning and data-intensive computing.