New Magnetic Tunnel Junction Neuron Enables Signed Spiking for Smarter AI
A 10nm MTJ device achieves signed spiking, boosting accuracy to 91% on CIFAR-10.
A new paper from researchers Huannan Zheng, Jingli Liu, and Kezhou Yang introduces a compact magnetic tunnel junction (MTJ) that acts as a signed spiking neuron. Unlike traditional spiking neurons which only output positive (excitatory) spikes, signed spiking neurons can also produce negative (inhibitory) spikes, enabling them to encode richer information. The key innovation is the use of orthogonal easy axes in the free and pinned layers of the MTJ, which allows the device to generate bipolar spike patterns and closely follow a signed leaky integrate-and-fire (LIF) equation.
The team built a representative device just 10nm × 45nm × 50nm in size. Using Landau–Lifshitz–Gilbert (LLG) simulations, they showed that proper free-layer dimensions make the device's magnetic dynamics match signed LIF behavior. When integrated into neural network evaluations using a fitted device-neuron model, the MTJ neuron achieved 91.06% accuracy on CIFAR-10 (image classification) and 77.40% on CIFAR10-DVS (event-based vision), retaining most of the accuracy of ideal mathematical signed LIF neurons. This work demonstrates that neuromorphic hardware can move beyond conventional binary spiking to handle both excitation and inhibition, potentially enabling more powerful AI accelerators with lower power consumption.
- Signed spiking neurons can encode both excitation and inhibition, delivering richer information than standard spiking neurons.
- The compact MTJ design uses 10nm × 45nm × 50nm dimensions with orthogonal easy axes to mimic signed LIF dynamics.
- Achieves 91.06% accuracy on CIFAR-10 and 77.40% on CIFAR10-DVS, nearly matching ideal software-based signed neurons.
Why It Matters
Enables compact, energy-efficient neuromorphic hardware that handles signed spiking, pushing AI closer to brain-like computation.