Optical Neural Network Inference Engine Using Photon Directing Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial neural networks (ANNs) and deep neural networks (DNNs) face delays and high power consumption due to the need for numerical calculations across multiple layers, which can be time-consuming and energy-intensive, especially as the complexity of calculations increases with the number of layers.
Innovation Solution
An all-optical transport implementation of an inference engine that performs inferences directly in the optical domain without converting to the computational or electrical domain, using photon directing devices with lenses formed according to learned weights to direct light and trigger optical sensors for classification, enabling near-instant inference and minimal power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional neural networks perform numerical calculations across multiple layers, then classification accuracy is achieved, but inference time increases and power consumption rises
Solution Approach 1:
The patent replaces the electrical/computational domain with the optical domain. Photon directing devices use optical components (lenses, mirrors) to direct light signals through the neural network layers, substituting electronic numerical calculations with optical signal routing. This enables parallel processing of calculations across multiple layers simultaneously, dramatically reducing inference time while maintaining classification accuracy.
Solution Approach 2:
The invention transitions from sequential electrical signal processing to parallel optical signal processing. By encoding neural network weights into the physical geometry of photon directing devices and using light propagation through multiple spatial layers simultaneously, the system performs computations in the optical dimension rather than sequential electrical cycles, achieving near-instant inference.
2Measurement precision
If conventional neural networks perform numerical calculations across multiple layers, then classification accuracy is achieved, but power consumption increases
Solution Approach 1:
The patent replaces energy-intensive electrical computations with passive optical signal routing. Photon directing devices use fixed optical components to direct light signals, eliminating the need for active electronic calculations during inference. This substitution dramatically reduces power consumption while maintaining the neural network's classification capabilities.
Solution Approach 2:
The optical neural network performs computations passively through the physical arrangement of photon directing devices and light propagation. The system uses the natural properties of light (reflection, refraction) to execute neural network operations without requiring external power for computational processing, only for initial optical signal generation and final detection.
3Adaptability or versatility
If neural network weights are stored in conventional computational systems, then model functionality is maintained, but security is compromised due to ease of reverse-engineering
Solution Approach 1:
The patent embeds neural network weights into the physical structure of photon directing devices rather than storing them in digital memory. The weights are encoded in the geometric properties (lens shapes, mirror angles) of optical components, making them physically embedded and difficult to extract or reverse-engineer while maintaining full model functionality.
Solution Approach 2:
The invention uses optical properties (analogous to color changes) to encode information. The physical characteristics of optical components (refractive indices, geometric shapes) serve as the encoding medium for weights, making the information intrinsic to the physical structure rather than stored as digital data that can be easily copied or reverse-engineered.
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 significantly reduces inference time and power consumption, allows for miniaturization of neural networks, and enhances security by hiding weights within the physical structure of the optical device, making it harder to reverse-engineer.
Implementation Method 1
An all-optical transport implementation of an inference engine that performs inferences directly in the optical domain without converting to the computational or electrical domain, using photon directing devices with lenses formed according to learned weights to direct light
Data Source
AI summary
Systems and methods are provided for an optical transport implementation of an inference engine capable of performing inferences in the optical domain. Examples include an optical device that includes photon directing devices disposed along an optical axis, each photon directing device corresponds to a layer of a trained machine learning model. Lenses are provided for each photon directing device, which are formed based on weights of a layer of the trained machine learning model corresponding to the respective photon directing device. The examples may also include optical sensors that correspond to inferences of the trained machine learning model, and the photon directing devices may be configured to receive light of an input and direct the light to one of the optical sensors according to the trained machine learning model.