Minimalist Optical Neural Computing: Optical Diffractive Neural Network by 2‐level Quantized Pixel‐Wise Optical Encoding DOI
Xianjin Liu, Ting Ma, Qiwen Bao

et al.

Laser & Photonics Review, Journal Year: 2025, Volume and Issue: unknown

Published: May 19, 2025

Abstract Diffractive optical neural networks (DONNs) offer high‐speed, energy‐efficient artificial intelligence (AI) computation but face challenges with misalignment and model‐to‐reality gaps. In this work, an ultra‐simplified DONN architecture based on a digital mirror device (DMD) camera, dubbed as m‐DONN, is introduced experimentally validated. Notably, within the m‐DONN framework, DMD acts both input layer solitary hidden layer, which trained 2‐level quantization, markedly differing from configuration found in traditional DONNs. This minimalism binarization of diffraction can result highly nonlinear correlation between encoded information output. A 10‐classification accuracy over 82% achieved MNIST dataset theoretical modeling experimental measurements, utilizing 10 000 test samples. Furthermore, employed to construct online reinforcement learning agent capable dynamically stabilizing virtual inverted pendulum. The inherent simplicity proposed computing system, coupled cost‐effective implementation using either active or passive key components, not only demonstrates extremely powerful yet simple neuromorphic setup also paves way for acceleration optoelectronic AI applications across variety scenarios.

Language: Английский

Minimalist Optical Neural Computing: Optical Diffractive Neural Network by 2‐level Quantized Pixel‐Wise Optical Encoding DOI
Xianjin Liu, Ting Ma, Qiwen Bao

et al.

Laser & Photonics Review, Journal Year: 2025, Volume and Issue: unknown

Published: May 19, 2025

Abstract Diffractive optical neural networks (DONNs) offer high‐speed, energy‐efficient artificial intelligence (AI) computation but face challenges with misalignment and model‐to‐reality gaps. In this work, an ultra‐simplified DONN architecture based on a digital mirror device (DMD) camera, dubbed as m‐DONN, is introduced experimentally validated. Notably, within the m‐DONN framework, DMD acts both input layer solitary hidden layer, which trained 2‐level quantization, markedly differing from configuration found in traditional DONNs. This minimalism binarization of diffraction can result highly nonlinear correlation between encoded information output. A 10‐classification accuracy over 82% achieved MNIST dataset theoretical modeling experimental measurements, utilizing 10 000 test samples. Furthermore, employed to construct online reinforcement learning agent capable dynamically stabilizing virtual inverted pendulum. The inherent simplicity proposed computing system, coupled cost‐effective implementation using either active or passive key components, not only demonstrates extremely powerful yet simple neuromorphic setup also paves way for acceleration optoelectronic AI applications across variety scenarios.

Language: Английский

Citations

0