Nanowire Neural Network Learns Just Like the Human Brain

Artificial neural networks could provide a solution to many modern problems including medical diagnosis, face identification systems and data mining. However, typical physical neural networks require a large amount of data to be stored in memory.

Researchers from the University of Sydney and the University of California at Los Angeles recently developed a novel neural network made from nanowires that is capable of learning and remembering “on the fly”. Their research is published in Nature Communications.

Artificial Neural Networks

Neural networks use artificial intelligence (AI) to model biological neurons in real-world learning and memory tasks. These systems have a wide range of applications and commonly work by processing large data sets and drawing conclusions through self-learning techniques. Machine learning (ML) capabilities are often used in speech recognition, financial applications and even for decision-making tasks in videogames.

Technology Networks’ Junior Science Editor Rhianna-lily Smith was joined by Ruomin Zhu in an exclusive interview to discuss their latest study on the development of the novel nanowire neural network. Credit: Technology Networks.

Many of the artificial neural networks used today need large amounts of energy to power, due to their need to learn through analyzing datasets. This requirement massively limits their application for worldwide use. Lead author Ruomin Zhu, a doctoral student at the University of Sydney Nano Institute and School of Physics, set out to develop a novel neural network that is far more energy efficient than its previous predecessors.

How can a nanowire neural network make a difference?

The neural network is made up of tiny nanowires that are only a billionth of a meter in diameter. The wires arrange themselves in a pattern to mimic the neural networks found in the brain and perform memory and learning tasks using simple algorithms. These algorithms respond to changes in electronic resistance at the points where the nanowires overlap, called resistive memory switching, to simulate what would happen at the synapses in the brain.

 

The researchers used the neural network to recognize and remember sequences of electrical pulses that corresponded to different images. They also tested the network in image recognition using images downloaded from the Modified National Institute of Standards and Technology (MNIST) database of handwritten digits, featuring a collection of 70,000 small grayscale images. The nanowire neural network scored 93.4% in correctly identifying test images.

The future of ML

This novel neural network offers the ability to not just remember simple tasks, but also perform tasks using dynamic data accessed online. “Our novel approach allows the nanowire neural network to learn and remember ‘on the fly’, sample by sample, extracting data online, thus avoiding heavy memory and energy usage,” said supervising researcher Professor Zdenka Kuncic.

 

Zhu also described how these results can offer huge advantages compared to the networks already used: “If the data is being streamed continuously, such as it would be from a sensor for instance, ML that relied on artificial neural networks would need to have the ability to adapt in real-time, which they are currently not optimized for.”

 

Reference: Zhu, Lilak, Loeffler, et al. Online dynamical learning and sequence memory with neuromorphic nanowire networks. Nat Comm. 2023. doi: 10.1038/s41467-023-42470-5

This article is a rework of a press release issued by the University of Sydney.