Advancements in Neural Networks Through Nanotechnology: Brain-Like Functions Achieved in Real-Time Learning

A collaborative research team from the University of Sydney and the University of California, Los Angeles (UCLA) has engineered a tangible neural network capable of real-time learning and memory functions, closely resembling the activity of neurons in the human brain. This pioneering work employs networks of nanowires to simulate the neural pathways found in the brain. The research holds considerable promise for creating efficient, energy-saving machine intelligence, especially in the context of real-time online learning.

Significant Milestone Reached in the Quest for Energy-Efficient, Agile Machine Intelligence

This marks the inaugural demonstration of a tangible neural network possessing the ability to learn and memorize dynamically, in a manner analogous to neural functions in the brain. The discovery paves the way for the development of energy-efficient machine intelligence capable of tackling complex learning and memory assignments in real-world scenarios.

The research findings, published today (November 1) in the journal Nature Communications, are the product of a joint effort between scientists at the University of Sydney and UCLA.

A high-resolution electron microscope captured images of the nanowire neural network, which self-organizes in a pattern similar to the children’s game ‘Pick Up Sticks’. The intersecting points of the nanowires function akin to the brain’s synapses, reacting to electrical currents. Credit goes to the University of Sydney for the imagery.

Ruomin Zhu, the study’s lead author and a doctoral student at the University of Sydney Nano Institute and School of Physics, stated, “The research validates the potential for employing nanowire networks in simulating brain-like learning and memory capabilities, which can be utilized for processing continuously streaming data.”

The Role of Nanowire Networks

Comprising minuscule wires with diameters in the order of billionths of a meter, nanowire networks self-organize into configurations that simulate neural networks, akin to those found in the human brain. These networks have the potential to execute specialized data processing tasks.

In terms of memory and learning, straightforward algorithms are used to react to alterations in electronic resistance at junctions where the nanowires intersect. This function, termed ‘resistive memory switching,’ occurs when electrical inputs meet changes in conductivity, akin to neural synapses in the human brain.

Research Conclusions and Prospects

For this study, the researchers employed the network to recognize and memorize sequences of electrical pulses representing images, modeling the human brain’s information processing methods.

Zdenka Kuncic, the supervising researcher and co-author, equated the memory task to the memorization of a phone number. The network also successfully performed image recognition tasks using the MNIST database of handwritten digits, a compilation of 70,000 small greyscale images commonly used in machine learning research.

Kuncic also highlighted, “Our prior research established the nanowire networks’ capacity for simple memory tasks. The current work expands on this by showing that tasks can be executed using real-time data accessed online.”

She further noted that this represents a substantial advancement, given the challenge of processing continuously changing data in large volumes. Traditional machine learning models store this data and then train on it, resulting in excessive energy consumption.

“The new methodology allows the nanowire neural network to learn and memorize dynamically, extracting data in real-time, thus circumventing extensive energy and memory usage,” she added.

Ruomin Zhu also pointed out additional benefits of processing data in real-time. “In cases where data is continuously streamed, such as from sensors, conventional machine learning models that rely on artificial neural networks are not currently optimized for real-time adaptability,” he explained.

The nanowire neural network in the study demonstrated benchmark machine learning performance, achieving a 93.4 percent accuracy rate in correctly identifying test images. The memory task involved recalling sequences comprising up to eight numbers. For both sets of tasks, data was streamed into the network to verify its capabilities for real-time learning and to demonstrate how memory function enhances learning.

Reference: “Online Dynamical Learning and Sequence Memory with Neuromorphic Nanowire Networks” by Zhu, Lilak, Loeffler, et al, published on 1 November 2023 in Nature Communications.
DOI: 10.1038/s41467-023-42470-5

Frequently Asked Questions (FAQs) about Nanowire Neural Network

More about Nanowire Neural Network

• Nature Communications Journal Article
• University of Sydney Nano Institute Research Publications
• UCLA Research on Neural Networks
• Overview of Nanotechnology in Neural Networks
• MNIST Database of Handwritten Digits
• Introduction to Resistive Memory Switching
• Basics of Physical Neural Networks