EPREVAEngineer Paper
COMPUTINGAnalysisElectronicsSemiconductorsComputing & AI for engineering

Oakland University simulates a spiking neuron from a transistor and magnetic tunnel junction

In LTspice simulations, a transistor paired with a magnetic tunnel junction reproduced neuron-like spiking, and a small network learned the XOR gate using spike timing. No hardware has been built.

Vivid and detailed close-up of a patterned silicon wafer with vibrant green and blue colors.
Photo: Nic Wood / Pexels

What the team proposed

Researchers at Oakland University in Michigan have designed an electronic neuron that pairs a conventional transistor with a nanoscale magnetic device, a magnetic tunnel junction (MTJ). The work was published in IEEE Transactions on Circuits and Systems I: Regular Papers. The team calls the approach "CMOS+X": it keeps standard CMOS electronics and adds a spintronic device alongside them.

The motivation is energy. Dr Steven Louis, the lead researcher, said in a press release that the human brain uses 20 watts of power, while the typical AI model needs about 500 times that for inference and another 50 times for each question asked. The team aims for AI that works locally rather than relying on large data centres.

How the magnetic neuron works

An MTJ contains magnetic layers separated by an extremely thin insulating barrier. One of those layers can change its orientation when current passes through it, because electrons carry spin, which the source describes as behaving somewhat like a tiny magnetic orientation. As current flows, those spins exert a small twisting force on the magnetisation, an effect called spin-transfer torque.

As the magnetisation shifts, the MTJ's electrical resistance changes. In memory chips that resistance change can represent a stored bit; here the researchers use it to help generate the short electrical pulses, or spikes, that biological neurons use to communicate.

The MTJ is paired with a single NMOS transistor, which controls the current through the magnetic device. As the magnetisation moves, the resistance changes and alters the voltage in the circuit; under the right conditions this produces a spike.

The simulated device reproduces several neuron-like behaviours: threshold spiking, latency, refractory periods, synaptic integration and inhibition. In conventional electronic neurons these behaviours often require complex circuits with many transistors; here, according to the researchers, they emerge from the magnetic dynamics of a single MTJ.

What the simulations showed

The results come from circuit simulations in LTspice, not from a fabricated device. The artificial neurons communicated through spike timing on nanosecond timescales, with the researchers reporting typical response latencies below five nanoseconds and a regime in which each neuron produces at most one spike per input pattern.

To test learning, the team connected the neurons into a small multilayer network and trained it on the XOR gate, a standard benchmark for nonlinear machine learning. The network began with synaptic weights that did not produce correct answers; during training it ran inference, compared actual output spike timing with the desired timing and adjusted the weights using a gradient-descent-based rule. After training, the network could reproduce the XOR function.

The researchers report that the network converged under the modelled conditions, demonstrating nonlinear processing, multilayer communication and weight adaptation entirely within the analog simulation, with training performed inside the simulated circuit rather than externally in software. The demonstration is small, and the source states that no one is suggesting a handful of magnetic neurons will replace a GPU.

Why timing carries the information

In a conventional digital system a signal is generally read as a logical one or zero. In a spiking system, a pulse that arrives early can mean something different from one that arrives later, so timing becomes part of the information processed.

The source describes the approach as potentially interesting for power-constrained, event-driven applications such as industrial sensors, robots, cameras, audio systems and wearable devices, where data arrives intermittently rather than as a constant stream. A spiking circuit can remain largely inactive until a spike arrives, rather than repeatedly calculating on data that has not changed.

Limits the team reports

The published results have so far been proven only in circuit simulations using LTspice. Fabricating the proposed neuron and demonstrating it in a laboratory will be a matter of time and funding, according to the source.

The source notes that real magnetic devices are messy: manufacturing variations can alter their behaviour, thermal fluctuations can disturb magnetisation, and temperature, resistance differences and self-heating may all affect how reliably spikes are produced. The next step is to establish how robust spike timing is under real-world conditions.

Some parts of the simulated system are also idealised, with certain synaptic and learning circuits represented by functional models rather than a complete transistor-level implementation. A practical neuromorphic processor would need much more than one working neuron: large numbers of connected devices, practical synapses, management of variation between components and training methods that do not turn the system back into a conventional digital computer.

Whether the approach translates into faster or more energy-efficient AI remains to be seen, the source states. MTJs already have an established manufacturing and research base through their use in MRAM and sensors; the open question is whether the same devices can become practical building blocks for computation.

What the researchers say

Louis said that putting enough of these elements together produces a result that begins to look less like a conventional processor and more like a collection of biological neurons. The source adds that future AI hardware may not need to calculate every part of a neural network explicitly, because some of the computation could happen through the behaviour of the materials themselves.

Technical fact box

  • Typical simulated neuron response latency: below five nanoseconds
  • Spikes per input pattern per neuron in the simulated regime: at most one
  • Benchmark learned by the simulated network: the XOR gate
  • Training method: gradient-descent-based rule adjusting synaptic weights
  • Simulation tool: LTspice
  • Circuit elements per neuron: one NMOS transistor and one magnetic tunnel junction
  • Brain power cited by lead researcher: 20 watts
  • Typical AI model inference power cited: about 500 times that of the brain

Sources

  1. A CMOS+X Spiking Neuron With On-Chip Machine Learning — IEEE Transactions on Circuits and Systems I: Regular Papers — Steven Louis, Matthew Blake Abramson, Hannah Bradley et al.
  2. This tiny magnet could turn ordinary transistor circuits into artificial neurons — Phys.org (Technology)

Discussion

Technical questions and corrections are welcome. Comments are signed and checked before they appear.

No comments yet.

Sign in to add a typed review comment. Reviews are signed with your name.

More analysis