Engineers at the University of Texas at Dallas have built a tiny brain-like computer that learns from experience using magnetic components, a breakthrough that could one day let phones and wearables run sophisticated artificial intelligence without draining batteries or relying on the cloud.
The prototype, made with magnetic tunnel junctions, mimics how human neurons and synapses strengthen or weaken connections as they learn. Unlike traditional AI systems that demand massive labeled datasets and power-hungry training, the new device taught itself to recognize simple patterns after just a handful of examples.
“Our work shows a potential new path for building brain-inspired computers that can learn on their own, said Joseph S. Friedman, the UT Dallas associate professor who led the research with colleagues from Everspin Technologies and Texas Instruments, according to a UTD news release. “Since neuromorphic computers do not need massive amounts of training computations, they could power smart devices without huge energy costs.”
In experiments, a grid of eight magnetic junctions correctly classified all 16 possible four-pixel images and, without any external guidance, specialized itself to recognize specific patterns in just nine learning cycles. Larger simulations using the standard MNIST handwriting dataset reached 90 percent accuracy while performing more than 600 trillion operations per second per watt — thousands of times more efficient than conventional graphics processors.
The key component is the spin-transfer torque magnetic tunnel junction, a nanoscale sandwich of two magnetic layers separated by an insulator. Its stable binary states provide reliable memory, while controlled random switching introduces the flexibility brains need to explore and settle on useful connections.
Friedman’s team applied Hebbian learning — the decades-old principle that “cells that fire together, wire together” — in hardware. When two artificial neurons were activated simultaneously, the magnetic link between them grew more conductive; when only one fired, the connection weakened.
“If one artificial neuron causes another artificial neuron to fire, the synapse connecting them becomes more conductive,” Friedman explained.
Traditional neuromorphic designs using memristors or phase-change materials often suffer from drifting resistance and reliability problems. The magnetic approach avoids those issues and can be manufactured with existing chip-making processes.
Researchers say the technology could eventually power autonomous robots, medical monitors that detect real-time health changes, or phones that adapt to their owners without ever sending data to distant servers.
The team’s research was recently published in the Nature journal Communications Engineering.
