Optical Neural Computing With Polarization-Encoded Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational efficiency of convolutional neural networks (CNNs) is a challenge for embedded systems due to power and bandwidth constraints, limiting their deployment in mobile devices and other compact systems, and existing optical computing systems are too large for portable applications.
Innovation Solution
A compact optical neural network (ONN) using an array of light emitting devices (LEDs) that control light beam polarization and intensity through electrical inputs, enabling efficient computation by emulating synaptic functions and storing weights in magnetization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If optical computing is used to improve computational efficiency, then processing speed and bandwidth are improved, but system volume becomes too large for portable applications
Solution Approach 1:
The system is divided into two functional segments: an optical computing segment for high-speed parallel processing operations, and an electronic control segment for weight storage and configuration. This segmentation allows each part to optimize its function while keeping the overall system compact.
Solution Approach 2:
The patent merges optical computing components with electronic control components into a single integrated system. The electronic weights are combined with optical modulators, and the optical processing unit is integrated with detection electronics, creating a hybrid architecture that reduces total system volume.
2Ease of manufacture
If physical phase masks are used as filters for convolution, then optical filtering is achieved, but the filters cannot be changed electrically
Solution Approach 1:
The patent replaces physical phase masks with electronically controllable modulators (such as liquid crystal modulators or electro-optic modulators). These devices use electrical signals to change optical properties, enabling dynamic reconfiguration of filtering operations without mechanical movement or physical replacement of components.
Solution Approach 2:
The system changes optical parameters (phase, amplitude, polarization) of light beams through electrical control of modulator parameters. By varying electrical input signals, the optical filtering characteristics can be dynamically adjusted, allowing the same hardware to perform different convolution operations.
3Measurement precision
If the number of parameters and connections in CNNs is increased to improve performance, then computational accuracy is improved, but power and memory requirements increase
Solution Approach 1:
The patent replaces electronic computation of weighted sums with optical physics-based operations. Light propagation through optical elements naturally performs matrix multiplication and convolution operations, transferring computational tasks from the electronic domain to the optical domain, thereby reducing dynamic power consumption.
Solution Approach 2:
The optical system performs computation passively using the properties of light propagation and interference. The optical components (lenses, modulators, waveguides) automatically perform mathematical operations on light beams without requiring active electronic processing, reducing the energy burden on the system.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The ONN provides efficient and compact neural network computation with fast computing speed and non-volatile storage, suitable for embedded systems, by leveraging the advantages of optical computing.
Implementation Method 1
a spin injector configured to inject spin-polarized carriers into the light emitting structure in response to a second input
Implementation Method 2
a light emitting structure configured to emit a light beam with a circular polarization rate (PC) corresponding to the first input and with an intensity corresponding to the second input
Data Source
AI summary
Disclosed is an optical apparatus for neural network computation. A light emitting device receives a first input and a second input, and emits a light beam with a circular polarization rate corresponding to the first input and with an intensity corresponding to the second input. A computing apparatus includes a plurality of light emitting devices, and may further include a light mixer and a light polarization detection system. A sum of products of the plurality of input data INPUTi (from second input) and respective weights Wi (from first input) assigned to the plurality of input data is thus calculated by the computing apparatus by means of optical operation. The computing apparatus can be configured to perform convolution calculation, and serves as a node in a hidden layer of the neural network. And a computing system for performing neural network computation is thus realized by means of optical computing.