Single-Ended EAM with Electrical Combining for Neuromorphic Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing photonic computing approaches for neuromorphic applications face scalability issues due to increased circuit complexity with multiple wavelengths and limitations in precision, especially in implementing large-scale neural networks, as they rely on wavelength-division-multiplexing schemes and optoelectronic conversions, which hinder the use of all-optical non-linear activation functions.
Innovation Solution
The development of coherent photonic circuit architectures that perform linear algebraic computations using optical splitters, amplitude modulators, phase shifters, and optical combiners, allowing for scalable and precise implementation of linear neurons and neural network layers, integrated with electronic circuitry for hybrid photonic-electronic computing systems, enabling flexible adaptation of neural network models and non-linear activation functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If wavelength-division-multiplexing schemes are used to encode neuron input signals onto different wavelengths, then the optical implementation of linear neurons is achieved, but the circuit complexity increases substantially with each added wavelength, limiting scalability
Solution Approach 1:
The patent transitions from wavelength-division-multiplexing (spectral dimension) to spatial-division-multiplexing (spatial dimension) by using multiple independent single-mode waveguides to carry different neuron input signals. This dimensional change eliminates the need for multiple wavelengths while maintaining the ability to encode multiple inputs, thereby reducing circuit complexity while preserving optical implementation precision.
2Adaptability or versatility
If optoelectronic conversion is used to enable addition or subtraction of weighted signals, then signed weights can be implemented, but the employment of all-optical non-linear activation functions is impeded
Solution Approach 1:
The patent extracts the sign-encoding function from the optical domain and relocates it to the electrical domain. By using electrical switches to apply sign values to photocurrents after optoelectronic conversion, the system maintains weight implementation flexibility while enabling subsequent all-optical non-linear activation functions to operate on the converted signals without electrical intervention during the activation process.
Solution Approach 2:
The patent introduces electrical switches as an intermediary component between the optoelectronic conversion stage and the non-linear activation stage. These switches enable sign-encoding operations without requiring the optical signals to remain in the optical domain throughout the entire computation, allowing flexible weight implementation while preserving the ability to use all-optical activation functions.
3Adaptability or versatility
If coherent electric-field addition is used for single-wavelength optical linear neurons, then signal errors accumulate along the cascade of MZIs, but precision high enough for large-scale practical applications cannot be achieved
Solution Approach 1:
The patent substitutes the coherent optical field addition mechanism (based on interference in MZIs) with direct electrical current addition after optoelectronic conversion. By converting optical signals to electrical photocurrents and performing the summation in the electrical domain, the system eliminates the error accumulation problem inherent in cascaded MZI coherent addition while maintaining the ability to implement optical linear neurons.
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
These architectures achieve scalable and precise linear algebraic computations, supporting high-density neural network implementations with low power and cost, while allowing for flexible functionality distribution between photonic and electronic domains, enhancing the performance of neuromorphic computing systems.
Implementation Method 1
an optical splitter configured to split the optical input signal into a plurality (m) of optical carrier signals
Implementation Method 2
a plurality of electro-optical modulators configured to modulate the plurality of optical carrier signals in accordance with a plurality of computational inputs and a plurality of computational weights to create a plurality of modulated optical signals
Implementation Method 3
a direct-detection photodetector configured to convert the modulated optical signal output into an electrical signal output
Data Source
AI summary
Disclosed here are systems, methods, and apparatuses for single-ended electro-absorption modulators (EAMs) with electrical combining. In particular, systems and methods are disclosed for performing optical encoding and multiplication operation for optical signal without applying a sign value. The optical output can be converted into a photocurrent input and the sign value can be applied to the photocurrent input on an electrical layer.