Unsupervised learning in hexagonal boron nitride memristor-based spiking neural networks

Area of Science:

• Materials Science
• Neuroscience
• Computer Engineering

Background:

• Resistive random access memory (RRAM) offers a path beyond von Neumann architectures by integrating processing and memory.
• Two-dimensional (2D) materials present superior non-volatile switching characteristics for RRAM applications compared to traditional oxides.
• Hexagonal boron nitride (h-BN) is a 2D material with potential for memristive applications.

Purpose of the Study:

• To investigate the electrical performance of 2D hexagonal boron nitride (h-BN) memristors for integration into spiking neural networks (SNNs).
• To evaluate the use of h-BN memristors as artificial synapses for unsupervised learning.
• To enhance SNN recognition rates using a novel STDP-based dropout technique.

Main Methods:

• Experimental characterization of h-BN memristor electrical behavior.
• Simulation of unsupervised learning in SNNs using experimental memristor data.
• Implementation and testing of a spike-timing-dependent-plasticity (STDP)-based dropout method.

Main Results:

• h-BN memristors exhibit viable performance as artificial synapses.
• Successful simulation of unsupervised learning for image classification on the MNIST dataset.
• Demonstrated enhancement in SNN recognition rates with the proposed STDP-based dropout technique.

Conclusions:

• 2D material-based memristors, specifically h-BN, are suitable for artificial synapses in SNNs.
• Unsupervised learning can be effectively implemented in SNNs using these memristors.
• Hardware-friendly online learning methods, like STDP-based dropout, can improve SNN performance.