Neuromodulation of Neuromorphic Circuits

Luka Ribar, Rodolphe Sepulchre

Abstract

We present a novel methodology to enable control of a neuromorphic circuit in close analogy with the physiological neuromodulation of a single neuron. The methodology is general in that it only relies on a parallel interconnection of elementary voltage-controlled current sources. In contrast to controlling a nonlinear circuit through the parameter tuning of a state-space model, our approach is purely input-output. The circuit elements are controlled and interconnected to shape the current-voltage characteristics (I-V curves) of the circuit in prescribed timescales. In turn, shaping those I-V curves determines the excitability properties of the circuit. We show that this methodology enables both robust and accurate control of the circuit behavior and resembles the biophysical mechanisms of neuromodulation. As a proof of concept, we simulate a SPICE model composed of MOSFET transconductance amplifiers operating in the weak inversion regime.

Introduction

In biological systems, neurons are not static processors. Through the process of neuromodulation, chemicals like dopamine or serotonin can alter a neuron's intrinsic properties, allowing it to transition between different functional states—such as moving from a resting state to rhythmic bursting. Replicating this flexibility in neuromorphic hardware is a significant challenge in brain-inspired engineering.

Traditional neuromorphic designs often suffer from "fragility." They typically rely on precise mathematical state-space models where switching between behaviors requires fine-tuning parameters to an impractical degree of precision. Furthermore, real-world silicon is plagued by "mismatch," where manufacturing variations mean no two transistors are identical, often breaking delicate mathematical models. This research proposes a design methodology that is both easily tunable (neuromodulable) and inherently resistant to the imperfections of physical hardware.

Methodology: I-V Curve Shaping

The proposed methodology shifts the focus from internal state variables to an input-output approach based on I-V curve shaping. The architecture treats the neuron as a parallel interconnection of elementary components, mimicking the way biological ion channels contribute to membrane potential.

The proposed parallel architecture separates the passive membrane properties from the active excitability elements.

Key components of this architecture include:

  1. Parallel Architecture: The circuit consists of a passive RC (resistor-capacitor) membrane in parallel with several "localized conductance" elements.
  2. Timescale Separation: These elements are categorized by their reaction speeds: Fast (milliseconds), Slow (tens of ms), and Ultra-slow (hundreds of ms). This separation is critical for defining complex behaviors like bursting.
  3. Hardware Implementation: The circuit is realized using MOSFET transconductance amplifiers operating in the weak inversion (subthreshold) regime. This allows the circuit to operate at ultra-low power levels, closely matching the energy efficiency of biological neurons.

Key Insights: Qualitative Control and Robustness

The core innovation lies in prioritizing the shape and monotonicity of the I-V curves over exact mathematical precision. By maintaining a specific "N-shape" in the fast timescale, the circuit can reliably produce excitability and spiking.

Shaping the I-V curves across fast and slow timescales determines the circuit's excitability and spiking thresholds.

This approach offers several advantages:

  • Neuromodulation as Gain Control: Modulating the circuit is simplified to changing the bias voltage (gain) of specific current sources. This directly maps to how biological modulators change the maximal conductance of ion channels.
  • Robustness to Mismatch: Because behavior is defined by the qualitative shape of the I-V curve, manufacturing defects that shift transistor thresholds can be "tuned out." Engineers can adjust control voltages to restore the target I-V curve and the desired neural behavior.
  • Rigid vs. Fragile Bursting: The authors demonstrate that their "double N-shape" method (overlapping two N-shaped regions in different timescales) creates "rigid" bursting that remains stable across a wide range of parameters, unlike more fragile mathematical models.

Results and Discussion

The methodology was validated through both numerical simulations in MATLAB and circuit-level SPICE simulations using a TSMC 0.35 $\mu$m process model. The SPICE simulations successfully demonstrated smooth transitions from resting to regular spiking and complex bursting simply by varying control voltages.

In the bursting regime, the simulated circuit consumed only 0.77 $\mu$W, highlighting its suitability for large-scale, energy-efficient neuromorphic chips. Furthermore, the researchers modeled a 10% variation in transistor parameters (threshold voltage, mobility, etc.) and showed that readjusting base voltages could perfectly restore nominal behavior, proving the architecture's resilience to process variation.

While the current work focuses on single-neuron dynamics, it lays the groundwork for future research into large-scale networks. A potential limitation is the reliance on clear timescale separation; if fast and slow scales overlap significantly, the qualitative shaping rules become more difficult to apply. Future work will explore these neuromodulable neurons in the context of Central Pattern Generators (CPGs) for robotics and physical hardware realization to test against real-world thermal noise.

Conclusion

By mimicking the biophysical mechanisms of neuromodulation through I-V curve shaping, this methodology provides a robust and intuitive framework for controlling neuromorphic circuits. It bridges the gap between complex biological behavior and the practical constraints of analog silicon, offering a path toward more flexible and resilient brain-inspired hardware.

Summary updated

Abstract

We present a novel methodology to enable control of a neuromorphic circuit in close analogy with the physiological neuromodulation of a single neuron. The methodology is general in that it only relies on a parallel interconnection of elementary voltage-controlled current sources. In contrast to controlling a nonlinear circuit through the parameter tuning of a state-space model, our approach is purely input-output. The circuit elements are controlled and interconnected to shape the current-voltage characteristics (I-V curves) of the circuit in prescribed timescales. In turn, shaping those I-V curves determines the excitability properties of the circuit. We show that this methodology enables both robust and accurate control of the circuit behavior and resembles the biophysical mechanisms of neuromodulation. As a proof of concept, we simulate a SPICE model composed of MOSFET transconductance amplifiers operating in the weak inversion regime.
BibTeX
@article{ribar2019neuromodulation,
  author = {Luka Ribar and Rodolphe Sepulchre},
  title = {{Neuromodulation of Neuromorphic Circuits}},
  journal = {IEEE Trans. Circuits Syst. I Regul. Pap.},
  year = {2019},
  url = {https://openreview.net/forum?id=vhsl7hDYts},
}
Figure