MZI Optical Circuit for Parallel Convolution Weight Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial neural networks (ANNs) face challenges with large operation depth and low efficiency when processing multiple groups of convolutional weights, particularly in convolution operations, which hinder the processing of multiple groups of convolution weights simultaneously.
Innovation Solution
The method involves constructing a convolution weight matrix, performing singular value decomposition (SVD) to obtain unitary and diagonal matrices, and determining Mach-Zehnder Interferometer (MZI) structures corresponding to these matrices, which are then connected to form an optical circuit that enables parallel convolution calculation processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional MZI-based convolution operation is used to process multiple groups of convolution weights, then the operation can be implemented with standard architecture, but the operation depth increases and operation efficiency decreases
Solution Approach 1:
The patent segments the convolution weight matrix into multiple sub-matrices, each processed by a separate MZI-based optical convolution unit. This segmentation allows parallel processing of multiple convolution weight groups simultaneously, reducing the overall operation depth while maintaining the ability to handle multiple convolution kernels. The segmentation transforms a sequential processing architecture into a parallel one, directly addressing the contradiction between operation efficiency and operation depth.
2Productivity
If multiple groups of convolution weights are processed sequentially in conventional ANN, then the implementation is straightforward, but the processing time increases and efficiency decreases
Solution Approach 1:
The patent merges multiple convolution operations into a single optical circuit by combining multiple weight matrices into a unified optical convolution system. The MZI-based optical circuit is configured to process multiple convolution weight groups simultaneously through parallel optical paths, merging what would traditionally be sequential operations into a concurrent process. This merging eliminates the time loss associated with sequential processing while maintaining straightforward implementation through the unified optical architecture.
3Adaptability or versatility
If conventional optical convolution is used for single convolution kernel, then the operation depth is manageable, but it cannot handle multiple groups of convolution weights efficiently
Solution Approach 1:
The patent creates a universal optical convolution circuit based on MZI structures that can handle multiple convolution weight groups through a single device configuration. The MZI-based optical convolution unit is designed with adjustable parameters that allow it to adapt to different convolution kernels and weight groups without requiring separate dedicated circuits for each. This multi-functionality enables the system to process multiple groups of convolution weights efficiently while keeping the operation depth manageable through the reusable, configurable optical architecture.
Data Source
AI summary
An optical circuit building method, an optical circuit, and an optical signal processing method and apparatus, the method comprising: constructing a convolution weight matrix corresponding to multiple groups of convolution weights (S11); performing singular value decomposition on the convolution weight matrix to obtain a first unitary matrix, a diagonal matrix and a second unitary matrix (S12); separately determining a first MZI structure corresponding to the first unitary matrix, a second MZI structure corresponding to the diagonal matrix and a third MZI structure corresponding to the second unitary matrix (S13); and connecting the first MZI structure, the second MZI structure and the third MZI structure to obtain an optical circuit (S14). In an optical circuit obtained by means of the foregoing, convolution calculation processing corresponding to multiple groups of convolution weights may be carried out at the same time, operation depth may be reduced, and operation efficiency may be improved.