Coherent Optical Neurons With Phase-Encoded Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical neuromorphic architectures face challenges in implementing positive and negative weight representations and all-optical non-sinusoidal activation functions while maintaining a reduced complexity, often requiring multiple wavelengths and electro-optical conversions.
Innovation Solution
A single-wavelength, coherent linear neuron stage using a multipath interferometer with electronically controlled phase shifters and amplitude modulators encodes weight signs in the optical phase, allowing for all-optical weighted summation and non-linear activation, implemented in photonic integrated circuits or fiber optics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If wavelength-division multiplexing schemes are used to encode input signals onto different wavelengths, then positive and negative weight representations are achieved, but the number of wavelengths increases with neuron fan-in and electro-optical conversion is required
Solution Approach 1:
The patent changes the encoding parameter from wavelength to optical phase. Instead of using different wavelengths to represent positive and negative weights, the invention uses phase modulation where the phase of the optical carrier encodes the sign of the weights, allowing all inputs to share a single wavelength while maintaining weight representation capability
Solution Approach 2:
The optical carrier serves multiple functions simultaneously: it carries all input signals at a single wavelength and encodes weight signs through phase modulation. This eliminates the need for separate wavelength channels for different weight signs, reducing the overall system complexity while maintaining adaptability
2Extent of automation
If two lasers at different wavelengths are used for each input to achieve positive and negative weights, then all-optical processing is enabled, but the system complexity and cost increase
Solution Approach 1:
The patent merges multiple wavelength channels into a single wavelength by using phase modulation instead of wavelength division. All input signals are processed optically using one laser at a single wavelength, with weight signs encoded in the phase domain, thereby maintaining all-optical processing while reducing the number of laser sources and wavelength channels
Solution Approach 2:
The invention uses the optical phase as a copy mechanism to represent weight signs without requiring separate physical optical paths or additional laser sources. The phase copy of the optical carrier encodes the sign information, enabling all-optical processing with a single laser
3Adaptability or versatility
If coherent neurons with multiple cascaded Mach-Zehnder interferometers are used for matrix multiplication, then positive/negative weight representations are achieved, but the spatial layout complexity increases
Solution Approach 1:
The patent extracts the weight sign encoding function from the complex spatial interferometer structure and implements it through phase modulation of a single optical carrier. This separates the weight representation function from the spatial layout, achieving matrix multiplication capability with reduced spatial complexity
Solution Approach 2:
The invention replaces the mechanical/spatial complexity of cascaded Mach-Zehnder interferometers with an electrical control system that modulates the phase of the optical carrier. The phase shifters, controlled by electrical signals, substitute for the complex optical path routing, reducing spatial layout complexity while maintaining computational capability
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 enables efficient, all-optical processing of weighted neuron inputs with positive/negative weights, supporting non-sinusoidal activation functions and reducing complexity, suitable for high-volume manufacturing and low-cost production.
Implementation Method 1
the modulated optical carrier signals are recombined at the outputs of the interferometer branches to generate an optical interference signal
Implementation Method 2
the signs of the neuron inputs and weights are imparted onto the phase of the optical carrier signal using one or more phase shifters
Implementation Method 3
the intensity of the optical interference signal is measured, e.g., by a photodetector, to determine an neuron output
Data Source
AI summary
Neuromorphic computing can employ coherent linear optical neurons implemented with multipath optical interferometers in which optical amplitude modulators and phase shifters impart neuron input signals and neuron weights onto optical carrier signals that are then summed at the interferometer output. Photodetectors at the interferometer output, optionally in conjunction with preceding nonlinear optical circuits, may implement a non-linear activation function and generate electronic neuron output signals that can be provided as input to other optical neurons to form an optical neural network.