An all-optical neuron with sigmoid activation function.

Area of Science:

• Photonics
• Optical Computing
• Artificial Neural Networks

Background:

• Traditional electronic neurons face limitations in speed and power consumption.
• Optical implementations offer potential for faster and more energy-efficient computation.
• Developing all-optical activation functions is crucial for photonic neural networks.

Purpose of the Study:

• To design and experimentally demonstrate an all-optical neuron.
• To implement a logistic sigmoid activation function using photonic components.
• To validate the neuron's performance in processing weighted optical signals.

Main Methods:

• Utilized a Semiconductor Optical Amplifier-Mach-Zehnder Interferometer (SOA-MZI) and a SOA-Cross-Gain-Modulation (XGM) gate to create the sigmoid function.
• Employed Wavelength-Division Multiplexing (WDM) for input signal multiplexing and weighting.
• Experimentally analyzed the transfer function and demonstrated neuron thresholding capabilities.

Main Results:

• Achieved an optical transfer function closely matching the logistic sigmoid.
• Demonstrated successful thresholding of a 100 picosecond pulse sequence.
• Showcased processing of 4 distinct weighted-and-summed power levels.

Conclusions:

• The presented all-optical neuron effectively implements a logistic sigmoid activation function.
• This work validates the feasibility of photonic components for neural network operations.
• The findings pave the way for high-speed, low-power optical computing architectures.