Fig. 1: Loose neuron array.
From: Optical neural network via loose neuron array and functional learning
We numerically verify loose neuron arrays with simulation. a The input neurons are point light sources, each with a 70 field of view, output neurons are energy gatherers, and LC neurons can attenuate incoherent light. These arrays are treated as black-box systems without gradient models and can be successfully trained using our FL paradigms. b Regular-2 array has two LC panels of a total of 2048 regularly aligned LC neurons. c Regular-3 array contains three layers of a total of 3096 LC neurons. d Normal-3 array is built on (b) with random jitters. e Uniform array has 3096 randomly distributed LC neurons. f The uniform array with random LC unions gains the highest performance, showing that data-driven training paradigms can outperform hand-craft hardware designs.
