Multilayer optical-electrical spiking neural network with sparse spike event for speech recognition based on a...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Photonics

Background:

• Spiking neural networks (SNNs) offer high power efficiency through event-driven computation and sparse activity.
• Optical platforms promise faster neural network processing but struggle with low-threshold nonlinear activation.
• Minimizing optical-electrical conversions is key for efficient optical neural network computation.

Purpose of the Study:

• To implement a multi-layer SNN with extremely sparse spike events for speech recognition on an optical-electrical platform.
• To demonstrate nonlinear activation within the optical domain using a novel laser.
• To assess the feasibility of photonic SNNs for complex tasks.

Main Methods:

• Developed a multi-layer SNN utilizing an extremely sparse spike event strategy (0.4 spikes/neuron on average).
• Employed a self-fabricated distributed feedback laser with a saturable absorber (DFB-SA) for optical nonlinear activation.
• Utilized the time-to-first spike encoding strategy for neural communication.

Main Results:

• Achieved a speech recognition accuracy of 90.5% on a benchmark dataset.
• Demonstrated significantly reduced spiking activity compared to other SNNs (approx. 4x fewer spikes).
• Successfully performed nonlinear computation within the optical domain.

Conclusions:

• The study validates the effectiveness of optical nonlinear activation for sparse SNNs in speech recognition.
• This work highlights the potential of photonic SNNs for efficient and complex computational tasks.
• Paves the way for advanced applications of SNNs in optical hardware.