Diffractive Optical Neural Networks Using Paired Signed-Signal Sensors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Opto-electronic sensors in diffractive optical neural networks are limited to detecting non-negative real numbers, limiting the range of realizable values and hindering real-time operation in complex optical tasks.
Innovation Solution
Implement a differential detection scheme using pairs of opto-electronic sensors per data class at the output plane, where one sensor captures the positive part and the other captures the negative part, and the final inference is based on the normalized difference between these sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional imaging-based machine vision systems use high pixel count and sensor array density to achieve high spatial resolution, then the performance of artificial neural networks is improved, but the requirements on memory size and computational power increase, inevitably hampering the effective frame-rate
Solution Approach 1:
The patent replaces traditional electronic neural networks with optical neural networks that use diffractive optical elements to perform computations. The optical system uses light propagation and diffraction to implement convolutional operations, eliminating the need for high-density sensor arrays and massive computational power while maintaining high spatial resolution and achieving real-time processing at the speed of light.
Solution Approach 2:
The patent changes the fundamental operating parameters from electronic signal processing to optical field manipulation. By using diffractive optical elements with specific transmission functions, the system transforms image processing tasks into optical physics problems, where the diffractive pattern directly encodes the neural network weights and the light field carries the input data, enabling parallel optical computation.
2Quantity of substance
If compressive sensing/sampling is used to overcome resource inefficiencies in conventional optical systems, then the number of detectors is reduced, but computationally demanding recovery algorithms partially hinder its application for real-time operation
Solution Approach 1:
The patent replaces computationally intensive electronic recovery algorithms with passive optical diffraction processes. The diffractive optical elements perform the compression and reconstruction operations in the optical domain using light propagation, eliminating the need for complex digital signal processing and enabling real-time operation with minimal computational overhead.
Solution Approach 2:
The patent introduces diffractive optical elements as intermediary components that bridge the input image and the final detection. These optical elements encode the neural network computations into their transmission function, allowing the system to perform both compression and classification in a single optical pass without requiring subsequent computational recovery steps.
3Device complexity
If diffractive optical neural networks use a single photo-detector per class to simplify the system, then the device complexity is reduced, but the detection range is limited to non-negative real numbers only
Solution Approach 1:
The patent segments the detection function by introducing separate positive and negative photo-detectors for each class. This segmentation allows the system to detect both positive and negative values by assigning different detectors to different detection tasks, thereby extending the detection range while maintaining the simplicity of the diffractive optical architecture.
Solution Approach 2:
The patent inverts the traditional approach by using the negative detector to capture the negative part of the output signal. Instead of trying to represent negative values through complex optical modulation, the system uses a simple inversion strategy where the negative detector directly measures the negative component, enabling the system to handle both positive and negative real numbers.
4Reliability
If additional diffractive layers are added to improve the generalization and inference performance of the network, then the accuracy is improved, but the device complexity and fabrication difficulty increase
Solution Approach 1:
The patent merges multiple diffractive layers into a single integrated optical system where each layer corresponds to a specific neural network layer. The diffractive elements are designed to work together in sequence, with each layer performing a specific computation and the output of one layer serving as the input to the next, thereby achieving deep neural network functionality while maintaining the simplicity of individual optical components.
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
This approach enhances blind testing accuracies to 98.54%, 90.54%, and 48.51% for MNIST, Fashion-MNIST, and grayscale CIFAR-10 datasets, respectively, approaching the performance of electronic deep neural networks.
Implementation Method 1
diffractive deep neural networks which are composed of successive diffractive optical layers (transmissive and/or reflective), trained and designed using deep learning methods in a computer, after which it is physically fabricated to all-optically perform statistical inference based on its trained task at hand
Implementation Method 2
With a single photo-detector assigned to each individual class of objects, Mengu et al. demonstrated a blind testing accuracy of 97.18% for all-optical classification of handwritten digits (MNIST database, where each digit was encoded in the amplitude channel of the input)
Data Source
AI summary
A diffractive optical neural network device includes a plurality of diffractive substrate layers arranged in an optical path. The substrate layers are formed with physical features across surfaces thereof that collectively define a trained mapping function between an optical input and an optical output. A plurality of groups of optical sensors are configured to sense and detect the optical output, wherein each group of optical sensors has at least one optical sensor configured to capture a positive signal from the optical output and at least one optical sensor configured to capture a negative signal from the optical output. Circuitry and/or computer software receives signals or data from the optical sensors and identifies a group of optical sensors in which a normalized differential signal calculated from the positive and negative optical sensors within each group is the largest or the smallest of among all the groups.