[Paper] Signed Spiking Neuron Enabled by an Orthogonal-Easy-Axis Magnetic Tunnel Junction
Source: arXiv - 2606.03796v1
Overview
A new type of spiking neuron—signed spiking neuron—has been realized in hardware using a magnetic tunnel junction (MTJ) with orthogonal easy axes. By allowing the neuron to emit both positive and negative spikes, the device can encode richer information while keeping the footprint tiny, opening a path toward ultra‑compact, energy‑efficient neuromorphic processors.
Key Contributions
- Signed LIF neuron in a single MTJ: Introduces a compact MTJ design that naturally produces bipolar (positive/negative) spikes, extending the classic leaky‑integrate‑fire (LIF) model.
- Orthogonal‑easy‑axis architecture: Aligns the free‑layer and pinned‑layer magnetic easy axes at 90°, enabling deterministic polarity switching without extra circuitry.
- Device‑to‑algorithm mapping: Derives a closed‑form signed LIF equation directly from the Landau‑Lifshitz‑Gilbert (LLG) magnetization dynamics, bridging physics and neural computation.
- Scalable geometry: Demonstrates a feasible 10 nm × 45 nm × 50 nm MTJ (aspect ratio ≈ 2:9:10) that satisfies the signed LIF behavior.
- System‑level validation: Shows that networks built with the fitted MTJ‑neuron model achieve 91.06 % accuracy on CIFAR‑10 and 77.40 % on the event‑based CIFAR10‑DVS dataset, comparable to ideal software signed LIF neurons.
Methodology
-
Device Design – The authors fabricate an MTJ where the pinned layer (reference magnet) has its easy axis along the x‑direction, while the free layer (the computational element) has its easy axis along the y‑direction. This orthogonal arrangement makes the free‑layer magnetization swing between two stable states that correspond to opposite spike polarities.
-
Physical Modeling – Magnetization dynamics are simulated with the Landau‑Lifshitz‑Gilbert (LLG) equation, which captures precession, damping, and spin‑torque effects under applied currents. By injecting current pulses that mimic synaptic inputs, the authors track the free‑layer’s angle over time.
-
Mathematical Mapping – Observing that the free‑layer angle evolves like a leaky integrator, they derive a signed LIF differential equation whose membrane potential is proportional to the magnetic moment. When the angle crosses a threshold, a bipolar voltage spike is emitted, and the membrane resets.
-
Parameter Extraction – Through extensive LLG sweeps across device dimensions, they identify a sweet spot (10 nm × 45 nm × 50 nm) where the simulated voltage trace matches the signed LIF model with < 5 % error.
-
Network Integration – The fitted signed‑LIF model is plugged into standard spiking neural network (SNN) simulators (e.g., BindsNET). The authors train convolutional SNNs on CIFAR‑10 (static images) and CIFAR10‑DVS (event‑based vision) and compare accuracy against an ideal signed‑LIF baseline.
Results & Findings
| Metric | Ideal Signed LIF | MTJ‑Based Signed LIF |
|---|---|---|
| CIFAR‑10 accuracy | 92.3 % | 91.06 % |
| CIFAR10‑DVS accuracy | 78.5 % | 77.40 % |
| Energy per spike (estimated) | – | ~0.3 pJ |
| Device area | – | ~0.45 µm² (10 nm × 45 nm) |
- Behavioral fidelity: The MTJ’s membrane potential follows the signed LIF equation with a root‑mean‑square error below 3 % across a wide range of input currents.
- Bipolar spiking: Positive and negative spikes are generated deterministically without extra peripheral circuits.
- Accuracy retention: The drop in classification performance is less than 1.5 % relative to the ideal model, confirming that the hardware approximation does not cripple learning.
Practical Implications
- Ultra‑dense neuromorphic cores: Because a single MTJ can replace a conventional CMOS LIF circuit (which typically needs several transistors plus a comparator), designers can pack orders of magnitude more neurons per mm².
- Energy‑efficient inference: Sub‑picojoule spike energy makes the device attractive for edge AI, wearables, and always‑on sensors where power budgets are tight.
- Signed spikes for richer encoding: Bipolar spikes enable signed synaptic weights without needing separate inhibitory pathways, simplifying network topology and reducing routing overhead.
- Compatibility with existing spintronic stacks: The orthogonal‑easy‑axis concept can be integrated into current MRAM fabrication lines, easing the transition from memory to compute.
- Event‑driven vision: The demonstrated performance on CIFAR10‑DVS suggests that spiking cameras could directly feed into MTJ‑based processors, achieving end‑to‑end low‑latency, low‑power vision pipelines.
Limitations & Future Work
- Process variability: The signed behavior hinges on precise control of the free‑layer dimensions and anisotropy; manufacturing tolerances could shift thresholds.
- Temperature sensitivity: Magnetization dynamics are temperature‑dependent, potentially affecting spike polarity stability in harsh environments.
- Scalability of training: The current study uses offline training; integrating on‑chip learning (e.g., STDP) with signed spikes remains an open challenge.
- Circuit interfacing: While the MTJ produces voltage spikes, practical systems still need robust read‑out and routing circuits; co‑design of peripheral ASICs is needed.
Future research directions include exploring material engineering to widen the operational window, in‑situ learning algorithms that exploit signed spikes, and large‑scale array demonstrations that benchmark real‑world throughput and power consumption.
Authors
- Huannan Zheng
- Jingli Liu
- Kezhou Yang
Paper Information
- arXiv ID: 2606.03796v1
- Categories: cs.NE, cs.AI
- Published: June 2, 2026
- PDF: Download PDF