Current-mode Silicon Neuron Achieves Robust Neuromodulation, Enabling Adaptive Neuromorphic Systems

Neuromorphic computing, which aims to build electronic systems that mimic the brain, promises adaptability, robustness and energy efficiency for applications ranging from edge computing to robotics. Loris Mendolia from University of Liège, Chenxi Wen from University of Zurich and ETH Zurich, and Elisabetta Chicca from University of Groningen, alongside Giacomo Indiveri, Rodolphe Sepulchre and Jean-Michel Redouté, now demonstrate a significant advance in this field with a novel silicon neuron capable of robust neuromodulation. This research introduces a current-mode neuron design that, crucially, adapts its behaviour in a biologically plausible manner, mirroring how real neurons respond to changing conditions. By achieving this adaptation with minimal circuit complexity and utilising standard manufacturing technologies, the team unlocks a pathway to more flexible, robust and scalable neuromorphic systems suitable for real-world deployment, offering a substantial step towards brain-inspired computing.

Analog VLSI Circuits for Neuromorphic Computation

This extensive collection of research materials details a project focused on building biologically-inspired neural circuits using analog Very Large Scale Integration (VLSI). The work centers on creating physical circuits that mimic the behaviour of biological neurons and synapses, rather than relying on software simulations. A key goal is to develop circuits that accurately model the dynamics of real neurons, considering specific neuron types known for their complex behaviour. The research focuses on Spiking Neural Networks (SNNs), where information is encoded in the precise timing of spikes, offering a different approach from traditional artificial neural networks.

The team explores neuromodulation, the process of altering neuron behaviour through chemical signals, and adaptive control of ion channel expression, aiming to create circuits that can learn and adapt over time. Furthermore, the project prioritizes low-power design, utilizing current-mode circuits and advanced fabrication technologies to minimize energy consumption, essential for large-scale neuromorphic systems. The inclusion of memristive devices suggests investigation into synapses with plasticity, the ability to change connection strength.

CMOS Neuron Design and Characterisation Setup

Scientists engineered a novel current-mode neuron using a 180nm CMOS process, fabricating an array of 16 neurons on a test chip measuring 750 × 90 μm². This design directly emulates biological neuron principles, focusing on robust neuromodulation through adaptable input response and spiking patterns. The team implemented a mixed-feedback architecture, enabling dynamic adjustment of neuron behaviour, and constructed a custom printed circuit board (PCB) to facilitate comprehensive characterization of the on-chip neurons. The experimental setup precisely controls the neuron’s internal state using 12 bias voltages generated by an external digital-to-analog converter.

A transmission gate multiplexer, managed by on-chip shift registers, selects one of three internal currents for off-chip readout, allowing for reliable measurement of very small currents. The team converted the output current into a voltage using a precision operational amplifier on the PCB, achieving adjustable gain as needed. Measurements were acquired using a high-resolution oscilloscope, ensuring accurate capture of neural traces. Researchers performed extensive measurements at room temperature, validating the neuron’s ability to produce and switch between different firing regimes, and observing consistent increases in firing frequency as input current increased.

Neuromodulation Emulated in Novel Neuron Circuit

Scientists have developed a novel current-mode neuron circuit that emulates key features of biological neurons, specifically their ability to adapt to changing conditions through neuromodulation. The research centers on a mixed-feedback architecture utilizing current-mode circuits, enabling compact, robust, and energy-efficient designs. This approach represents the internal dynamics of a neuron with three distinct timescales, achieved through a combination of low-pass filters and nonlinear feedback loops. The core of the circuit utilizes differential-pair integrators and current-mode sigmoid blocks, incorporating a biologically inspired positive feedback inactivation mechanism to precisely control neuron behaviour.

Experiments on a 180nm CMOS implementation confirmed the theoretical predictions, demonstrating robust spiking and bursting behaviour. Measurements reveal an energy efficiency ranging from 40 to 200 picojoules per spike, showcasing the circuit’s potential for ultra-low power applications. Further testing demonstrated the circuit’s resilience to temperature variations, maintaining stable operation across a range of 5 to 45 degrees Celsius. The inclusion of slow positive feedback significantly enhances the expressivity of the model, enabling the generation of complex patterns like bursting.

Robust Neuromodulation in CMOS Neuron Circuits

This research presents a novel current-mode neuron circuit designed to emulate the adaptive behaviour of biological neurons through neuromodulation, a process where neurons alter their responses based on context. The team successfully implemented a circuit that exhibits robust and tunable neuromodulation with minimal complexity, compatible with standard manufacturing technologies. Mathematical modelling underpinned the circuit’s design, allowing for precise analysis and tuning of neuron behaviour, which was then validated through both simulations and experimental results using a low-power prototype. The fabricated 180nm CMOS implementation demonstrates the circuit’s ability to spike, burst, and undergo real-time neuromodulation across a range of temperatures, confirming its robustness and adaptability to variations in operating conditions. Importantly, the neuron operates with ultra-low power consumption, consuming only 40 to 200 pJ per spike, making it suitable for integration into large-scale neuromorphic systems. Ongoing work focuses on creating networks of these neurons with programmable connections and neuromodulatory pathways, aiming to demonstrate adaptive signal processing and sensorimotor control applications, and ultimately advance the field of energy-efficient analog computing.