Abstract
The Hindmarsh-Rose (HR) neuron model is built to describe the neuron electrical activities. Due to the polynomial nonlinearities, multipliers are required to implement the HR neuron model in analog. In order to avoid the multipliers, this brief presents a novel smooth nonlinear fitting scheme. We first construct two nonlinear fitting functions using the composite hyperbolic tangent functions and then implement an analog multiplierless circuit for the two-dimensional (2D) or three-dimensional (3D) HR neuron model. To exhibit the nonlinear fitting effects, numerical simulations and hardware experiments for the fitted HR neuron model are provided successively. The results show that the fitted HR neuron model with analog multiplierless circuit can display different operation patterns of resting, periodic spiking, and periodic/chaotic bursting, entirely behaving like the original HR neuron model. The analog multiplierless circuit has the advantage of low implementation cost and thereby it might be suitable for the hardware implementation of large-scale neural networks.
Introduction
Neuromorphic engineering aims to replicate the biological brain's efficiency and complex dynamics using hardware. A cornerstone of this field is the Hindmarsh-Rose (HR) model, a mathematical framework capable of simulating a vast range of neuronal activities. However, the HR model's reliance on cubic ($x^3$) and quadratic ($x^2$) nonlinearities presents a significant hurdle for analog circuit design.
Traditionally, implementing these polynomial terms requires analog multipliers. These components are often expensive, consume substantial silicon area, and increase overall circuit complexity. While digital alternatives like FPGAs often employ "piecewise linear" approximations to bypass multipliers, these methods are non-smooth and non-differentiable. Such inaccuracies can compromise the fine-grained dynamical behaviors, particularly chaotic states, that make the HR model scientifically valuable.
Methodology: Smooth Nonlinear Fitting
To address the limitations of traditional multipliers and non-smooth approximations, this research proposes a smooth nonlinear fitting scheme. This approach replaces complex polynomial math with composite hyperbolic tangent ($\tanh$) functions.
Function Construction
The researchers designed two primary fitting functions, $H_1(x)$ and $H_2(x)$, using combinations of $\tanh$ terms. Specifically, $H_1(x)$ utilizes three $\tanh$ terms to approximate the cubic nonlinearity, while $H_2(x)$ uses two terms for the quadratic nonlinearity. As shown in Figure 1, these functions achieve high fidelity to the original polynomials.
Circuit Realization and Prototyping
The choice of $\tanh$ functions is physically motivated: the current-voltage relationship of a transistor differential pair naturally follows a $\tanh$ curve. This allows the functions to be implemented using standard analog components—operational amplifiers (TL082CP), transistors (MPS2222), and resistors—eliminating the need for dedicated multiplier chips. A physical prototype was constructed on a Printed Circuit Board (PCB), using a potentiometer to adjust the steady current ($I$) and sweep through various neuronal firing states.
Results and Discussion
The performance of the fitted model and the physical circuit was rigorously benchmarked against numerical simulations of the original HR equations.
Behavioral Replication
The fitted model successfully replicated four distinct neuronal operation patterns: resting, periodic bursting, chaotic bursting, and periodic spiking. These patterns were consistent across both numerical simulations and physical hardware experiments.
Bifurcation Consistency
A critical finding is the model's ability to reproduce the Inter-Spike Interval (ISI) bifurcation diagrams. This demonstrates that the fitted neuron accurately mimics the transitions between different firing regimes as the input stimulus changes, preserving the underlying dynamical structure of the original model.
Hardware Precision and Scalability
The hardware implementation proved robust, with a Mean Absolute Percentage Error (MAPE) of only 2.4% between theoretical and measured resistance values. This high precision, combined with the low cost of multiplierless design, suggests that this scheme is well-suited for large-scale neural network implementations.
Conclusion and Future Work
By leveraging the inherent physics of transistors to create smooth $\tanh$-based approximations, the researchers have developed a cost-effective, multiplierless analog circuit for the HR neuron model. While this study focuses on a single neuron, future work will explore the implementation of coupled networks and assess power/area trade-offs at scale. Additionally, the sensitivity of these analog circuits to environmental factors like temperature remains an area for further investigation.
Reference: Cai, J., Bao, H., Xu, Q., et al. (2021). Smooth Nonlinear Fitting Scheme for Analog Multiplierless Implementation of Hindmarsh-Rose Neuron Model. Published in Nonlinear Dynamics. DOI: 10.1007/s11071-021-06453-9
Summary updated