Mott memristor based stochastic neurons for probabilistic computing

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computational Neuroscience

Background:

• Probabilistic spiking in biological systems enhances learning and Bayesian inference.
• Stochasticity in nanoscale devices offers potential benefits for neuromorphic systems.

Purpose of the Study:

• To develop a stochastic leaky integrate and fire (LIF) neuron utilizing Mott memristor dynamics.
• To demonstrate the neuron's capability for biological neural dynamics and probabilistic computation.
• To integrate the neuron into advanced neural network architectures for learning and inference tasks.

Main Methods:

• Fabrication of a stochastic LIF neuron incorporating a Mott memristor.
• Integration of the neuron into a population-coded spiking neural network and a spiking restricted Boltzmann machine (sRBM).
• Evaluation of the sRBM's accuracy for probabilistic learning and inference.

Main Results:

• The developed LIF neuron exhibits biological neural dynamics.
• The integrated sRBM achieved a high accuracy of 87.13%, comparable to software implementations.
• The design eliminates the need for external noise sources, unlike CMOS-based probabilistic neurons.

Conclusions:

• The Mott memristor-based stochastic LIF neuron enables energy-efficient and compact neuromorphic systems.
• The proposed neuron effectively implements probabilistic learning and inference in spiking neural networks.
• This approach offers a promising pathway for advanced, low-power neuromorphic computing applications.