Artificial Neuron Circuits Mimic Brain’s Processing with Targeted Electrical Signals

Researchers are increasingly investigating bio-inspired computing architectures to overcome the limitations of conventional systems. Wendy Otieno, Alex Gabbitas, and Debi Pattnaik, from the Department of Physics at Loughborough University, alongside Pavel Borisov, Sergey Savel’ev, and Alexander G. Balanov, demonstrate the stochastic dynamics of diffusive memristor blocks designed to mimic synaptic convergence found in natural neural circuits. Their work addresses a critical gap in understanding how these complex neural blocks function and how to effectively implement them in hardware. By modelling and experimentally verifying input voltage combinations that enable targeted activation within a simplified network of three spiking artificial neurons, the team reveals key statistical characteristics of spiking patterns and provides insights towards building universal functional blocks for future neuromorphic systems.

Diffusive memristors enable simplified synaptic convergence in a three-neuron neuromorphic block with reduced complexity

Researchers have demonstrated a functional three-neuron block constructed from diffusive memristors, paving the way for more efficient neuromorphic computing systems. This work centres on synaptic convergence, a common feature of natural neural circuits, and successfully implements a simplified artificial circuit utilising diffusive memristors to mimic this process.
Numerical modelling and subsequent experiments have revealed specific input voltage combinations capable of triggering targeted spiking activity within the three-neuron configuration. The study details the construction and analysis of a neuromorphic block comprising three diffusive memristors, each paired with a load resistor and capacitor.

External voltages applied to the first two neurons influence the spiking behaviour of the third, post-synaptic neuron, effectively modelling a simplified sensory system. Through careful control of input voltages, researchers were able to selectively activate individual neurons or the entire block, demonstrating a level of control crucial for computational applications.

The team’s analysis extends to the statistical characteristics of these spiking patterns, interpreting them from a computational standpoint and validating the numerical simulations with experimental measurements. Diffusive memristors, unlike traditional memory devices, exhibit volatile switching behaviour, mimicking the transient nature of biological synapses.

These devices consist of a metal filament embedded within a solid electrolyte, forming a conductive path when voltage is applied and dissolving upon its removal. This inherent volatility, coupled with the devices’ scalability, energy efficiency, and high switching speed, makes them ideal candidates for neuromorphic architectures.

The ability of these memristors to replicate complex biological behaviours, such as synaptic plasticity and spike-timing-dependent plasticity, further enhances their suitability for brain-inspired computing. The research highlights the potential for utilising the inherent stochasticity of diffusive memristors, similar to the noise present in biological neurons, for computational purposes.

While neuromorphic technologies continue to evolve, understanding how simple neuron blocks contribute to overall computing performance remains critical. The successful creation of this three-neuron block, with a demonstrated ability to control spiking activity, represents a significant advance in neuromorphic computing hardware development0.1, however, the creation of a functional three-neuron block represents a significant advance in neuromorphic computing hardware development.

Fabrication of diffusive memristor crossbar junctions and Pearson-Anson oscillator emulation demonstrate neuromorphic computing potential

Diffusive memristors formed the core of an experimental investigation into artificial neural circuits. Three identically-deposited diffusive memristors, designated RM1, RM2, and RM3, were fabricated using magnetron sputtering deposition and UV photolithography techniques to create a three-neuron block.

Bottom electrodes consisting of 5nm titanium and 45nm gold were patterned onto SiO2/Si wafers, utilising photolithography to define 200μm long, 10μm wide paddle-shaped electrodes. A 50nm switching layer of SiOx:Ag was then deposited via co-sputtering at rates of 0.6Å/s for SiO2 and 0.1Å/s for Ag. Finally, top electrodes of 5nm titanium and 120nm gold were deposited perpendicularly, forming a 10μm x 10μm crossbar junction of SiOx:Ag between the patterned electrodes.

Each memristor was connected in parallel with a 1 nF capacitor and in series with a load resistor to emulate Pearson-Anson oscillators, encouraging spiking behaviour. These parallel capacitors were significantly larger than the internal capacitance of approximately 24 pF exhibited by the memristors when subjected to voltages ≤1.0V, effectively negating any parasitic effects.

To initiate self-oscillating behaviour, 10s duration voltage pulses were simultaneously applied to the series load resistors R1 and R2 connected to memristors RM1 and RM2. RM3 received input solely from the outputs of RM1 and RM2, and was monitored directly using a PicoScope 5443D oscilloscope. Three sequential pulses were applied, separated by 10s ‘OFF’ periods, to improve statistical reliability and minimise the impact of stochasticity.

Load resistor values were set at R1 = R3 = 55 kΩ and R2 = 60 kΩ. Input voltages V1 and V2 were systematically varied from 0V to 3.6V in 0.2V increments, creating a comprehensive range of input combinations. Spiking behaviour was determined by establishing that voltage oscillations exceeded the baseline noise level observed during OFF periods by at least 10%, a criterion that had to be met in at least two of the three ON periods to classify a device as spiking.

Current measurements, recorded with Keithley 7410 digital multimeters across 1 kΩ monitoring resistors, captured spiking in RM1 and RM2. Numerical simulations employed a charge transport model, successfully used in previous studies of diffusive memristors, representing each memristor with stochastic differential equations describing the motion of a metallic cluster between a conductive filament and a contact.

Diffusive memristor circuit exhibits targeted spiking via synaptic convergence and plasticity

Researchers demonstrated the successful operation of a three-neuron block constructed from diffusive memristors, achieving targeted activation of spiking patterns through precise input voltage combinations. Numerical modelling and subsequent experiments confirmed a strong correlation between simulated and measured spiking behaviours, validating the circuit’s functionality.

The study focused on synaptic convergence within this simplified artificial neural circuit, contributing to the development of functional building blocks for neuromorphic computing systems. Analysis of spiking patterns revealed statistical characteristics that provide insight into the circuit’s operational dynamics.

The research employed a charge transport model, successfully reproducing experimental results and allowing for detailed investigation of the memristors’ behaviour. Parameters were carefully selected to create non-identical neurons, each with slightly differing spiking thresholds and timescales, enhancing the circuit’s complexity and responsiveness.

To quantify spiking regularity, the team calculated the coefficient of variation of inter-spike intervals, CV1, and the local variability of inter-spike intervals, CV2. These metrics were used to assess the stochastic properties of the spiking activity. Probability density estimation of the bivariate data, using CV1 and CV2, was performed to analyse the correlation between these two characteristics.

The dimensionless model equations used in the simulations incorporated parameters such as qi/L = 0.2, k = 0.2, Ch = 0.18, T0 = 1.1, and λ = 0.13, alongside initial conditions of xi(0) = −0.95, Ti(0) = 30, and VMi(0) = 3.1. These values, combined with resistance values of R1 = 560, R2 = 610, and R3 = 560, facilitated the observation of targeted spiking behaviour. The Euler-Maruyama method, with a time step of h = 0.001, was used for numerical integration of the stochastic model equations.

Diffusive memristor circuits replicate cortical neuron spiking dynamics and logic functions with high energy efficiency

Researchers have demonstrated a functional three-neuron block based on diffusive memristors that mimics synaptic convergence found in biological neural circuits. This artificial neural circuit integrates inputs and produces outputs through the coordinated spiking of interconnected artificial neurons. Numerical modelling and experimental validation confirm the ability to activate specific spiking patterns by manipulating input voltages.

The circuit’s behaviour was analysed to reveal distinct voltage combinations that reliably trigger different neuron configurations, enabling functions such as signal comparison and Boolean logic operations. Statistical characteristics of the spiking patterns, quantified by coefficients of variation, align with those observed in cortical neurons, suggesting potential for biomimicry in neuromorphic systems.

While the initial theoretical model accurately predicted circuit behaviour, discrepancies with experimental results were noted, attributed to the dynamic nature of the conducting filament within the memristors. Addressing this non-stationarity through material engineering, current control, or alternative memristor technologies could further enhance circuit stability and reproducibility.

Key Number: 0.35, 0.7, representing the range of coefficients of variation observed, however, the creation of a functional three-neuron block represents a significant advance in neuromorphic computing hardware development. Future work may focus on stabilising the memristive filament to improve performance and exploring the integration of these functional blocks into larger, more complex neuromorphic architectures.