Core Concepts
Optical random neural networks can achieve high performance through genetically programmable random projection kernels.
Abstract
Machine learning tools, especially artificial neural networks, are essential for various applications.
Optical computing offers parallelism and fundamental operations with passive components.
Genetically programmable optical neural networks improve accuracy by optimizing random projection kernels.
Experimental setup includes a laser source, spatial light modulator, lenses, scattering medium, and camera.
Genetic algorithm optimizes the orientation of the diffuser to enhance classification accuracy.
Results show significant improvements in classification accuracy for various datasets using this approach.
Stats
By genetically programming the orientation of the scattering medium, accuracies improved 7-22% for machine learning tasks.
Test accuracies increased from 70% to 77%, 78% to 93%, and 58% to 80% for different datasets.
Quotes
"We demonstrate a genetically programmable yet simple optical neural network to achieve high performances with optical random projection."
"Our novel method is based on finding an optimum random projection kernel to map information optically for designated machine learning applications."