A collaborative study between the University of Sydney and the University of California at Los Angeles has unveiled a physical neural network’s capability to learn and memorize in a dynamic manner, closely resembling the brain’s neuronal functions. Published in Nature Communications, this advancement heralds the potential to cultivate efficient, low-energy machine intelligence tailored for intricate, real-time learning and memory operations.
The Brain-Inspired Nanowire Networks
Constructed from ultra-thin wires with a diameter in the billionth-of-a-metre range, these nanowire networks spontaneously form patterns evocative of the popular children’s game ‘Pick Up Sticks’. The formations generated imitate neural networks akin to the human brain, designated for specific data processing functions.
Resistive Memory Switching: A Synapse Mimicry
The method through which memory and learning activities are attained revolves around elementary algorithms that detect alterations in electronic resistance at overlapping nanowire junctions. Termed ‘resistive memory switching’, this phenomenon emerges when electrical inputs undergo conductivity shifts, drawing parallels with synaptic events in human brains. In the current study, these networks were deployed to discern and memorize sequences of electrical impulses linked to visual stimuli, reflecting human brain information processing methodologies.
Online Learning and Memory with Nanowires
The investigative team demonstrated the network’s proficiency in recognizing sequences analogous to phone number recall. Moreover, they employed the network for a quintessential image recognition task, retrieving images from the MNIST database, a compendium of 70,000 small greyscale images utilized for machine learning. It was previously validated that these nanowire networks could retain basic tasks. The present research amplifies these insights by confirming the ability to conduct tasks with dynamic online data. Notably, accomplishing online learning is intricate with voluminous and perpetually evolving data sets. A conventional strategy would mandate data storage followed by machine learning model training, incurring substantial energy costs. This innovative strategy empowers the nanowire neural network to learn and memorize dynamically, sourcing data in real-time, thereby circumventing excessive energy and memory consumption.
Benefits of Online Data Processing
Continuous data streaming, as one would expect from sensors, necessitates real-time adaptability, an area where current artificial neural networks fall short. In this study, the nanowire neural network exhibited a notable machine learning aptitude, registering a 93.4 percent success rate in accurate image identification. Additionally, the network demonstrated the capability to remember sequences comprising up to eight digits. Data was continuously fed into the system, illustrating its potential for real-time learning, underscored by memory-augmented enhancements.