Resonant Tunneling Diodes Replicate Neuronal Spiking for Opto-electronic Neuromorphic Computing

Inspired by the human brain, neuromorphic computing seeks to create artificial intelligence systems that are faster and more energy-efficient than conventional computers, and researchers are increasingly turning to novel hardware to achieve this goal. Dafydd Owen-Newns of the University of Exeter, along with colleagues, demonstrates a significant step forward by building artificial neurons and networks using resonant tunnelling diodes, tiny components that mimic the spiking behaviour of biological neurons. This work showcases how these light-activated “spiking neurons” can process information, detect rapid changes in data, and even function as a rewritable optical memory, offering a promising pathway towards ultra-fast, energy-efficient computing systems and advanced artificial intelligence applications. By constructing and testing both simple circuits and more complex neural networks, the team highlights the potential of photonic neuromorphic computing to overcome the limitations of current technology.

Photonic techniques for neuromorphic computing hardware are attracting increasing research interest, thanks to their potential for ultra-high bandwidths, low crosstalk and high parallelism. Approaches based upon resonant tunnelling diodes (RTDs) have recently gained attention as potential building blocks for next-generation light-enabled neuromorphic hardware, stemming from their capacity to replicate key neuronal behaviours such as excitable spiking and refractoriness, added to their potential for high-speed operation and low energy consumption. Consequently, researchers are actively investigating RTD-based photonic systems to create more efficient and biologically inspired computing architectures

Hybrid Photonics and Resonant Tunneling Diodes

This research details the construction of neuromorphic computing systems using a hybrid photonic-electronic approach. The core innovation lies in creating ultrafast and compact leaky integrate-and-fire (LIF) circuits, fundamental building blocks of artificial neural networks, by combining the speed of photonics with the compact size and functionality of resonant tunneling diodes (RTDs). The research addresses the limitations of traditional computers in tasks like pattern recognition and machine learning, which are well-suited to the brain’s efficient, parallel processing. LIF circuits are essential for simulating neurons, but building them efficiently, fast, small, and with low power consumption, presents a significant challenge.

The proposed solution leverages the strengths of both technologies. Photons facilitate fast signal transmission and parallel processing, while RTDs provide compact, nonlinear functionality for integration and firing. RTDs exhibit negative differential resistance, making them ideal for implementing the nonlinear dynamics required in LIF circuits. Researchers designed a specific circuit architecture using RTDs and photonic components to emulate the behaviour of a biological neuron, implementing the leak electronically and using photonic components to accelerate input summation. The resulting hybrid circuits demonstrated significantly faster operation compared to purely electronic LIF circuits, achieving a very compact circuit footprint.

Circuit parameters, such as leak rate and firing threshold, can be tuned to control the neuron’s behaviour, and the circuits successfully emulated key neural functions, including integration of inputs and generation of spikes, or action potentials. The research demonstrates the potential for large-scale integration due to the compact size and relatively low power consumption. The team modelled chaotic dynamics using the Mackey-Glass equation, potentially for more complex neural models, and tested their circuits using the Fisher’s Iris dataset, a standard benchmark for machine learning algorithms, demonstrating their ability to perform pattern recognition. This work contributes to the development of spiking neural networks (SNNs), considered a more biologically realistic and energy-efficient approach to neural computation, and positions itself within the broader field of neuromorphic computing, highlighting the advantages of its hybrid photonic-electronic approach compared to other technologies like memristors, CMOS circuits, and purely optical systems. In essence, this research presents a promising pathway towards building faster, smaller, and more energy-efficient neuromorphic computing systems by intelligently combining the strengths of photonics and electronics, with the use of RTDs as key nonlinear elements and the demonstration of functional LIF circuits representing significant contributions to the field.

Resonant Tunnel Diodes Emulate Biological Neurons

Researchers have developed a novel approach to neuromorphic computing using resonant tunneling diodes (RTDs), devices that mimic the behaviour of biological neurons with remarkable efficiency. This technology promises significant advancements in artificial intelligence and event-based systems, offering potential for ultra-fast processing and low energy consumption. The team successfully demonstrated that RTDs can function as optical-electronic spiking neurons, capable of processing information encoded in both light and electrical signals. The core of this innovation lies in the RTD’s ability to generate spikes, brief bursts of electrical activity, when stimulated by either light or voltage changes.

By carefully controlling the bias of the RTD, researchers can tune its sensitivity, effectively creating a threshold for spike generation. This allows the device to detect rapid changes in incoming signals, such as the rising edges of a time-series data stream, with exceptional speed and precision. Simulations and experiments confirm that these RTD neurons can operate at frequencies exceeding gigahertz, while consuming only picowatts of power per spike, significantly lower than many conventional computing systems. Furthermore, the team demonstrated the potential of these RTD neurons in more complex architectures.

They constructed a two-layer feedforward spiking neural network, utilizing RTDs as the non-linear processing nodes, and achieved excellent performance in classifying complex datasets. They also created a multi-neuron system functioning as an adjustable optical spiking memory, capable of storing and recalling patterns of spikes with tunable duration. This suggests the possibility of creating neuromorphic systems that can not only process information but also store and retrieve it in a biologically inspired manner. A key achievement is the ability to combine photonic and electronic signals for processing, allowing the RTD neuron to respond to changes in both light intensity and electrical voltage.

This dual modulation capability enables the detection of edge features in time-series data by encoding the raw signal into an optical pulse and using a delayed electrical signal to dynamically adjust the spiking threshold. This approach allows for event-based detection, where the system only responds to significant changes in the input signal, further enhancing efficiency and speed. The results demonstrate a pathway towards creating compact, energy-efficient, and high-speed neuromorphic systems with broad applications in areas such as real-time data analysis, pattern recognition, and adaptive robotics.

Optical Spiking Neurons Detect Features, Classify Data

This work demonstrates several novel applications of photonic-electronic spiking resonant tunnelling diode neurons for efficient neuromorphic data processing. Researchers successfully implemented these neurons in single-device systems, and integrated them into larger network architectures, showcasing their versatility. The investigations reveal that these RTD neurons can be activated by short pulses of optical power, with the required pulse amplitude linked to the applied bias voltage, creating a spike activation threshold suitable for various neuromorphic tasks. The team demonstrated rising edge-feature detection in time-series data using dual modulation of voltage bias and optical input, and achieved high classification accuracies, 96.
GHz rates. Furthermore, they created neuromorphic spiking memory systems, including a single RTD neuron capable of indefinite spike retention and a larger network of coupled devices. The authors acknowledge that the relationship between memory depth and attenuation within the spiking memory system was non-linear, and suggest that future research could explore optimising this relationship and investigating the scalability of these systems for more complex applications.