Photon convolution accelerator based on mode multiplexing

Abstract

The invention discloses a photon convolution accelerator based on mode multiplexing, and belongs to the field of optical calculation. Light output by a laser enters different corresponding waveguides after being split, the light in the different waveguides is written into picture information through an information writing unit, and then the product of signals and weights in corresponding convolution kernels is achieved through amplitude control. And the light after amplitude control is multiplexed to a multimode waveguide through a mode multiplexer to form different modes to be detected by a high-speed multimode photoelectric detector to output a convolution sum signal, and finally convolution layer processing on picture information is realized. According to the method, information is processed in parallel by using orthogonality of different modes in the waveguide, the capability that a traditional electronic chip cannot achieve parallel processing is broken through, convolutional layer processing in the convolutional neural network is accelerated, a brand new thought and a framework are provided for optical calculation processing of the convolutional neural network, and the method has a wide application prospect in the field of optical calculation and has a wide application prospect in the field of optical calculation. And the blank in the related technical field is filled.

Description

technical field [0001] The invention belongs to the field of optical computing, and more particularly, relates to a photon convolution accelerator based on mode multiplexing. Background technique [0002] In the past 50 years, the semiconductor technology has gradually matured. From the initial 180nm process technology to the current 5nm advanced technology, and the 3nm process currently under development, the development of integrated circuits has been rapid, and computing can be rapidly improved, but this state cannot be maintained forever. As the development continues, as the size of the transistor gradually decreases, the power consumption and heat dissipation of the chip are gradually revealed, which may become a stumbling block on the road of Moore's Law. [0003] The explosive growth of artificial intelligence requires a rapid increase in computing power, and traditional electronic computing methods are gradually reaching their performance limits and falling behind th...

AI Technical Summary

Problems solved by technology

These networks use multi-layer interconnected artificial neural networks to perform complex mathematical operations. Due to the high computational complexity and the need to train large data sets, the shortcomings of traditional serial computing electronic computers are gradually revealed.

Images