Compact photonic spiking neuron with inherent stochasticity based on phase-change material for probabilistic...

Area of Science:

• Photonics
• Neuromorphic Computing
• Materials Science

Background:

• Probabilistic models in photonic neural networks are promising for Bayesian inference.
• Existing photonic neurons lack intrinsic stochasticity, complicating designs.
• On-chip integration of probabilistic computing requires efficient stochastic neurons.

Purpose of the Study:

• To report the first compact on-chip photonic spiking neuron with inherent stochasticity.
• To utilize a novel phase-change material for probabilistic photonic computing.
• To demonstrate the neuron's capability for Bayesian inference and its robustness.

Main Methods:

• Developed a photonic spiking neuron using a novel phase-change material (SbTe9).
• Exploited intrinsic fluctuations in the material's melting point for stochasticity.
• Integrated the neuron into a system for Bayesian inference and tested its tolerance to variations and noise.

Main Results:

• Achieved stable and tunable probabilistic firing behaviors.
• Demonstrated 98.67% accuracy in breast cell diagnosis using Bayesian inference with uncertainty quantification.
• Showcased remarkable tolerance to hardware synaptic variations and input noise compared to deterministic neurons.

Conclusions:

• The novel stochastic photonic neuron enables low-complexity, high-performance on-chip photonic neuromorphic computing.
• Phase-change materials offer a transformative pathway for advanced neuromorphic systems.
• This work paves the way for large-scale, efficient probabilistic photonic computing architectures.