Tensorized Optical Neural Network PDE Solver
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Solving partial differential equations (PDEs) using traditional numerical methods is computationally intensive and requires significant resources, especially when dealing with high-dimensional or inverse problems.
Innovation Solution
The use of optical neural networks (ONNs) as PDE solvers without back-propagation (BP), utilizing tensorized optical neural network (TONN) inference accelerators fabricated on photonic chips, which leverage ultra-low-power wavelength-parallel photonic tensor cores to reduce resource requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional numerical methods are used to solve PDEs, then solution accuracy can be achieved, but computational resources and time consumption increase significantly
Solution Approach 1:
The patent replaces traditional electronic numerical computation systems with optical computing systems. Optical neural networks use light propagation and interference to perform computations, substituting the mechanical/electrical operations of traditional computers with optical processes that can solve PDEs faster and with lower power consumption while maintaining solution accuracy
Solution Approach 2:
The patent introduces optical weight matrices and photonic tensor cores as intermediary components that enable efficient computation. These optical intermediaries transform the PDE solving process by using light-based matrix operations to approximate solutions, bridging the gap between traditional numerical methods and optical computing
2Measurement precision
If traditional numerical methods are used to solve high-dimensional PDEs, then solution completeness can be achieved, but system resource requirements increase significantly
Solution Approach 1:
The patent segments the large-scale weight matrices into smaller photonic tensor cores that can be implemented on optical chips. This segmentation allows high-dimensional PDE problems to be broken down into manageable optical computation units, reducing the physical resource requirements while maintaining solution completeness through parallel optical processing
Solution Approach 2:
The patent transitions from electronic dimension scaling to optical dimension scaling by using spatial light modulation and wavelength multiplexing. This allows the system to handle high-dimensional PDEs by utilizing the spatial and spectral dimensions of light, effectively increasing computational capacity without proportionally increasing physical resource consumption
3Use of energy by moving object
If optical neural networks are used to solve PDEs, then resource consumption is reduced, but training complexity without back-propagation increases
Solution Approach 1:
The patent extracts and eliminates the back-propagation training mechanism from the optical neural network. By removing this complex iterative optimization process, the system reduces training complexity while maintaining the energy efficiency benefits of optical computing through direct forward-propagation-based solving of PDEs
Data Source
AI summary
A system for back-propagation free training of a tensor-compressed optical neural network (TONN) of a TONN inference accelerator. The system performs an iterative training process. In a given iteration of the process, a model input generator generates encode input data and encoded parameters, and the TONN inference accelerator is forward evaluated based on the input data and parameters. A loss evaluator receives an output of the TONN inference accelerator and evaluates the loss of the TONN based on the received output. A zeroth-order optimizer estimates a gradient of the loss. Then, in a next iteration of the iterative training process, the encoded parameters are updated based on the gradient of the loss as estimated in the previous iteration.