Ultrafast neural sampling with spiking nanolasers.

Area of Science:

• Neuromorphic computing
• Photonics
• Artificial intelligence

Background:

• Optical neuromorphic systems offer advantages in bandwidth, power efficiency, and latency over digital electronics.
• Photonic crystal nanolasers exhibiting excitable behavior have been recently demonstrated, emitting optical pulses (spikes) on a nanosecond timescale.

Purpose of the Study:

• To theoretically demonstrate the use of networks of photonic spiking neurons for Bayesian inference.
• To derive translation rules from conventional sampling networks to photonic spiking networks.
• To evaluate the potential of these systems for generative tasks and compare their performance to existing neuromorphic systems.

Main Methods:

• Theoretical derivation of translation rules from Boltzmann machines to photonic spiking networks.
• Demonstration of functionality across various generative tasks using simulated photonic spiking neuron networks.
• Estimation of processing speed and power consumption for the proposed optical neuromorphic system.

Main Results:

• Networks of photonic spiking neurons can perform Bayesian inference by sampling from learned probability distributions.
• The proposed system demonstrates functionality across a range of generative tasks.
• Expected improvements of several orders of magnitude in processing speed and power consumption compared to current state-of-the-art neuromorphic systems.

Conclusions:

• Photonic spiking neuron networks represent a promising avenue for efficient Bayesian inference in neuromorphic computing.
• The developed translation rules enable the implementation of complex AI tasks on optical hardware.
• This work highlights the potential for significant advancements in optical neuromorphic systems.