Diffraction Neural Network and Implementation Method Using Pyramid Structure Diffraction Layer to Complement Light

Abstract

The invention discloses a diffraction neural network and an implementation method using a pyramid structure diffraction layer to supplement light, wherein the diffraction neural network divides the diffraction layer in the network into two groups, which are respectively the diffraction layer of the supplementary light area and the diffraction layer of the diffraction area; Change the size of the diffractive layer in the supplementary light area to make it present a pyramid structure, so that a part of the input light can bypass several diffractive layers in the supplementary light area; the diffraction area is between the supplementary light area and the detection layer, and the diffraction layers in the diffractive area are all standard size. The method of the invention firstly designs the basic scheme of the diffraction neural network, confirms the relationship between its network performance and the number of layers, and then designs a supplementary light scheme according to task requirements and network characteristics. The invention builds more layers of networks by increasing the brightness, improves the network performance, increases the selection space of diffractive layer materials and detectors, and enhances the environmental adaptability and adjustment space of the network in practical applications; there is no introduction of additional light sources and devices , small size, low energy consumption, easy to install and apply.

Description

technical field [0001] The invention relates to a diffractive neural network using a pyramid-structured diffractive layer to supplement light and a realization method thereof. By adjusting the optical diffractive layer structure, the received brightness of the detection layer is improved, thereby increasing the depth of the network. It belongs to the field of combining deep learning and physical optics, and mainly relates to light Physical implementation of diffraction and neural networks. It has broad application prospects in various neural network application systems and schemes based on optical diffraction devices. Background technique [0002] Traditional computers are designed based on the von Neumann scheme and are built from logic units composed of transistors. The development speed of integrated circuits following Moore's Law in the past few decades is gradually facing technical bottlenecks. Photons do not generate heat, and are not affected by electromagnetic indu...

AI Technical Summary

Problems solved by technology

[0004] Energy loss caused by diffractive layers is an important issue for diffractive neural networks, which limits the number of achievable layers of optical diffractive neural networks

Images