Hybrid Optical Neural Network Training via Linear Layer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical neural networks (ONNs) face challenges in training due to experimental imperfections that are hard to accurately model, leading to suboptimal performance compared to in silico training.
Innovation Solution
A method of training neural networks where a mathematically linear stage of the forward propagation is performed optically, allowing for direct incorporation of optical signal updates into the weight matrices, thereby enhancing hybrid training and reducing the reality gap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If in silico training is used to train optical neural networks, then training can be performed digitally with standard algorithms, but the trained network performs worse than expected due to experimental imperfections in the physical system
Solution Approach 1:
The patent introduces an optical linear layer as an intermediary component between the digital training algorithm and the physical optical system. This layer is trained optically using real optical signals, allowing the network to adapt to actual experimental imperfections while maintaining the benefits of digital backpropagation for weight updates. The optical linear layer serves as a bridge that translates digital training objectives into the physical optical domain, closing the reality gap between simulation and experiment.
2Reliability
If simulated noise is incorporated into in silico training to account for experimental imperfections, then the reality gap may be narrowed, but the approach is suboptimal because it does not incorporate the specific pattern of imperfections present in a given ONN
Solution Approach 1:
The patent changes the training parameter from purely digital weight updates to a hybrid approach where the optical linear layer's parameters (weights) are updated through real optical signal propagation. This allows the network to learn the specific pattern of imperfections present in the actual optical system by experiencing them directly during training, rather than relying on generic simulated noise. The parameter update process incorporates real optical measurements, enabling adaptation to the specific hardware characteristics.
3Reliability
If multiple times repetition of optical testing and re-training is performed as in previous approaches, then corrections can be made to account for measured performance, but the process is time-consuming and the physics of later layers remains unaccounted for
Solution Approach 1:
The patent performs preliminary optical training of the optical linear layer during the main training process, rather than requiring separate iterative correction passes after digital training. By incorporating optical signal propagation and weight updates into the standard backpropagation workflow from the beginning, the network adapts to optical imperfections in real-time without needing multiple sequential correction passes. This preliminary action eliminates the need for repeated optical testing and re-training cycles.
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 improves network performance compared to in silico training, especially under static noise conditions, and allows for faster training while closing the reality gap by directly incorporating physical imperfections into the training process.
Implementation Method 1
a mathematically linear stage of the forward propagation is performed optically
Implementation Method 2
Thanks to the superposition and coherence properties of light, neurons in ONNs can be naturally connected via interference or diffraction in different settings
Implementation Method 3
Thanks to the superposition and coherence properties of light, neurons in ONNs can be naturally connected via interference or diffraction in different settings
Implementation Method 4
the neuron activation function can be physically implemented with a large variety of nonlinear optical effects
Data Source
AI summary
Methods and apparatus for training a neural network are disclosed. In one arrangement, a method comprises performing forward propagation of information through the neural network. An error backpropagation is performed to update parameters defining the neural network. A mathematically linear stage of the forward propagation is performed optically.