22nm Resonate-and-Fire Neuron Enables Linear State-Space-Models for Efficient Machine Learning

Oscillatory dynamics are increasingly important in machine learning, particularly for processing long sequences of data, and researchers are now exploring how to mimic the behaviour of biological neurons to improve these systems. Angqi Liu, Filippo Moro, and Sebastian Billaudelle, along with colleagues at the Institute of Neuroinformatics, University of Zurich and ETH Zurich, have developed a new electronic neuron that combines the benefits of oscillatory circuits with the efficiency of spike-based communication. This innovative ‘resonate-and-fire’ neuron, built using advanced silicon technology, operates with a linear design, making it more robust and predictable than previous models. The team demonstrates that this neuron not only functions reliably despite variations in manufacturing and operating conditions, but also performs well in practical applications like keyword spotting, establishing it as a promising building block for energy-efficient hardware.

Temporal processing relies on capturing oscillatory behaviour, and Resonate-and-Fire (RAF) neurons offer a spiking framework with strong expressivity and sparse event-based communication. This work introduces a new RAF neuron, built in 22nm Fully-Depleted Silicon-on-Insulator technology, that aligns with State Space Model (SSM) principles while retaining the efficiency of spike-based signalling. The team analysed the neuron’s dynamics, linearity, and resilience to variations in manufacturing, voltage, and temperature, and evaluated its power, performance, and area trade-offs. These characteristics demonstrate the neuron’s capabilities in temporal processing applications.

Oscillatory State Space Models in Analog Circuits

This research centres around implementing a neural circuit inspired by State-Space Models, known for their ability to process sequential data efficiently. The authors present a novel analog circuit, the RAF neuron, designed to mimic the dynamics of these oscillatory SSMs with low power consumption. A key design principle is maintaining linearity in the circuit’s behaviour, crucial for accurately representing the mathematical operations within the SSM. The RAF circuit features a two-state coupling mechanism, representing the core dynamics of the SSM, achieved through transconductance amplifiers and switched-capacitor circuits. System-level simulations, using the Spiking Heidelberg Digits dataset for keyword spotting, demonstrate the circuit’s performance, even when realistic hardware constraints, such as non-ideal linearity and device mismatch, are introduced.

Resonate-and-Fire Neuron Achieves Low Power Consumption

Scientists have developed a new resonate-and-fire (RAF) neuron, built using 22 nanometer Fully-Depleted Silicon-on-Insulator technology, that embodies the principles of State-Space Models while retaining the energy efficiency of spike-based communication. This work establishes RAF neurons as robust, energy-efficient computational primitives for neuromorphic hardware, bridging the gap between continuous-time oscillatory dynamics and discrete spiking networks. The circuit utilizes capacitors to store internal state voltages, and switched-capacitor leakage paths achieve precise control over the decay of these states. Analysis of the transconductance amplifiers demonstrates the linearity of the coupling between the states, a critical feature for accurate signal processing. Researchers mapped the characteristics of the circuit into a system-level simulation, demonstrating its effectiveness in a keyword-spotting task, and validating the co-design approach where circuit characteristics are accounted for in system-level simulations. This achievement simplifies the SSM equation and distills the core principles of SSMs into a biologically inspired, energy-efficient spiking framework, paving the way for advanced neuromorphic systems capable of processing long temporal sequences with high efficiency and robustness.

Resonate-and-Fire Circuit Enables Robust Keyword Spotting

This work demonstrates a new resonate-and-fire (RAF) neuron circuit, built using advanced silicon technology, that aligns with recent oscillatory models employed in machine learning, such as State-Space Models. The team successfully implemented a linear two-state coupling within the circuit, achieving both high linearity and time constants suitable for real-time signal processing. Through detailed analysis and modeling, researchers established the circuit’s resilience to variations in manufacturing processes, voltage, and temperature, confirming its robustness for practical applications. Evaluations within a system-level simulation, specifically a keyword-spotting task, revealed that the circuit’s inherent non-idealities do not significantly hinder performance, demonstrating the efficacy of a co-design approach. The results establish RAF neurons as energy-efficient primitives for hardware implementation, offering a promising pathway for advancing neuromorphic computing.