Self-Rectifying Dynamic Memristor Circuits for Periodic LIF Refractory Period Emulation and TTFS/Rate Signal Encoding

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) offer computational efficiency advantages over traditional artificial neural networks.
• Accurate emulation of biological neuron behavior, including the refractory period, is key to advancing SNN capabilities.
• Dynamic memristors present a promising avenue for hardware implementation of neural functions.

Purpose of the Study:

• To investigate the use of Ta/IGZO/TaO_x/Pt dynamic memristors with peripheral circuits to emulate leaky integrate-and-fire neuron behavior.
• To incorporate a refractory period into the memristor-based neuron model for enhanced biological accuracy and precise signal timing.
• To configure the memristor as an encoder for converting external signals into voltage pulse sequences using rate and time-to-first-spike (TTFS) coding.

Main Methods:

• Emulation of leaky integrate-and-fire neuron dynamics using dynamic memristors with nonlinear I-V hysteresis.
• Integration of a refractory period mechanism into the memristor circuit to prevent over-activation.
• Configuration of the memristor as a signal encoder utilizing rate coding and TTFS coding methods.

Main Results:

• The dynamic memristor successfully emulated neuron functions: integration, leakage, and firing, with an incorporated refractory period.
• The memristor encoder demonstrated efficient signal processing, achieving TTFS within 21-62 ms and operating at frequencies of 2500-9500 Hz.
• Experimental results confirmed enhanced Spiking Neural Network performance using the memristor-based neuron model.

Conclusions:

• Dynamic memristors, when combined with peripheral circuits, can accurately mimic biological neuron behavior, including the crucial refractory period.
• This memristor-based approach significantly improves Spiking Neural Network performance for real-time and temporal signal processing.
• The study highlights the potential of dynamic memristors in advancing neuromorphic computing systems for greater efficiency and biological plausibility.