pith. sign in

arxiv: 2606.03796 · v1 · pith:3SCEYQEJnew · submitted 2026-06-02 · 💻 cs.NE · cs.AI

Signed Spiking Neuron Enabled by an Orthogonal-Easy-Axis Magnetic Tunnel Junction

Huannan Zheng , Jingli Liu , Kezhou Yang This is my paper

Pith reviewed 2026-06-28 07:26 UTC · model grok-4.3

classification 💻 cs.NE cs.AI
keywords magnetic tunnel junctionsigned spiking neuronorthogonal easy axisleaky integrate-and-fireneuromorphic hardwarespiking neural networkCIFAR-10
0
0 comments X

The pith

An MTJ device with orthogonal easy axes in its layers functions as a signed leaky integrate-and-fire neuron.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to establish that a compact magnetic tunnel junction with orthogonal easy axes between free and pinned layers can generate bipolar spikes and follow signed leaky integrate-and-fire dynamics. Simulations confirm that suitable dimensions make the magnetic response match the signed LIF equation. A sympathetic reader would care because signed neurons transmit richer information than binary spiking ones, opening paths to more capable neuromorphic hardware. The device model then supports network accuracies close to those of ideal signed LIF units on image datasets.

Core claim

With orthogonal easy axes in the free and pinned layers, the magnetic tunnel junction enables bipolar spike generation and maps its magnetic-moment dynamics to signed leaky integrate-and-fire membrane-potential evolution. Landau-Lifshitz-Gilbert simulations show that proper free-layer dimensions allow the device response to follow a signed LIF equation. A representative design of 10 nm x 45 nm x 50 nm corresponds to an aspect ratio of about 2:9:10. Network evaluations using the fitted device-neuron model achieve 91.06% on CIFAR-10 and 77.40% on CIFAR10-DVS, retaining most of the accuracy of ideal signed LIF neurons.

What carries the argument

Orthogonal-easy-axis magnetic tunnel junction that maps magnetic-moment dynamics to signed LIF membrane-potential evolution

If this is right

  • The device supports bipolar spike generation for richer information than standard MTJ neurons.
  • Chosen dimensions make the magnetic response match the signed LIF equation.
  • Networks built on the device model reach 91.06 percent accuracy on CIFAR-10.
  • Performance stays close to that of ideal signed LIF neurons across tested datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The design could be combined with other MTJ-based components to reduce overall system power in spiking networks.
  • Variations in layer thickness might be explored to tune the leak time constant for different tasks.
  • Real-device measurements would reveal whether the signed mapping holds under operating temperatures typical of integrated circuits.

Load-bearing premise

The Landau-Lifshitz-Gilbert simulations with chosen free-layer dimensions accurately represent real-device behavior without unmodeled effects such as thermal fluctuations or fabrication imperfections.

What would settle it

Fabricate the 10 nm by 45 nm by 50 nm orthogonal-easy-axis MTJ and record its output under stepped input currents to test whether the effective membrane potential follows the signed LIF differential equation.

Figures

Figures reproduced from arXiv: 2606.03796 by Huannan Zheng, Jingli Liu, Kezhou Yang.

Figure 1
Figure 1. Figure 1: Schematic structure of the proposed MTJ-based device Table I. DEVICE AND SIMULATION PARAMETERS Parameter Symbol Value Unit Short axis of free layer 𝑊 10 nm Long axis of free layer 𝐿 45 nm Free-layer thickness 𝑡 50 nm Saturation magnetization [12] 𝑀𝑠 1.15 × 106 A/m Gilbert damping coefficient [13] α𝐺 0.01 - Spin efficiency [14] η 0.5 - Temperature 𝑇 300 K [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: a) Behavior of the unit magnetic moment of the free layer under current pulses with opposite polarities b) Evolution of the free-layer magnetic moment component along the pinned-layer easy-axis direction under an input pulse sequence c) Circuit design for bipolar spike generation based on the proposed device d) Equivalent membrane-potential dynamics with positive and negative thresholds [PITH_FULL_IMAGE:f… view at source ↗
Figure 3
Figure 3. Figure 3: a) Comparison of the state evolution among the ideal LIF neuron, the proposed device, and the conventional MTJ-based design. b) Magnetic-moment trajectories under short-axis input and long-axis input, showing the different dynamic behaviors induced by the two configurations. The physical origin of this difference is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
read the original abstract

Signed spiking neurons carry richer information than standard spiking neurons. This work proposes a compact magnetic tunnel junction (MTJ)-based neuron for signed leaky integrate-and-fire (LIF) operation. With orthogonal easy axes in the free and pinned layers, the device enables bipolar spike generation and maps magnetic-moment dynamics to signed LIF membrane-potential evolution. Landau--Lifshitz--Gilbert simulations show that proper free-layer dimensions allow the device response to follow a signed LIF equation. A representative design of 10 nm x 45 nm x 50 nm corresponds to an aspect ratio of about 2:9:10. Network evaluations using the fitted device-neuron model achieve 91.06% on CIFAR-10 and 77.40% on CIFAR10-DVS, retaining most of the accuracy of ideal signed LIF neurons.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes an MTJ-based signed spiking neuron with orthogonal easy axes between free and pinned layers to enable bipolar spike generation. It claims that magnetic-moment dynamics under LLG evolution can be mapped to signed LIF membrane-potential behavior for appropriate free-layer dimensions (example: 10 nm × 45 nm × 50 nm). Deterministic LLG simulations are used to identify dimensions where the response follows a signed LIF equation; a fitted device-neuron model is then evaluated in networks, yielding 91.06% on CIFAR-10 and 77.40% on CIFAR10-DVS while retaining most accuracy of ideal signed LIF neurons.

Significance. If the signed-LIF mapping is robust, the work would supply a compact, CMOS-compatible hardware primitive for signed spiking neurons, increasing information capacity per neuron relative to standard binary spikes. The concrete network accuracies provide a quantitative benchmark of end-to-end utility. Credit is due for the explicit dimension example and the end-to-end network evaluation, though both rest on a fitted model derived from deterministic simulations.

major comments (2)
  1. [Abstract / LLG simulations] Abstract and LLG simulation section: the central claim that “proper free-layer dimensions allow the device response to follow a signed LIF equation” rests entirely on deterministic LLG runs. No Langevin thermal-noise term is included, yet at the stated 10 nm scale the kT fluctuation strength is comparable to anisotropy energy and can destroy both the linear leak regime and the bipolar threshold crossing required for the signed-LIF mapping. This omission is load-bearing for the device-to-neuron equivalence.
  2. [Abstract] Abstract: no simulation parameters (damping, saturation magnetization, temperature, integration timestep), no goodness-of-fit metric (e.g., RMS error or R² to the ideal signed-LIF ODE), and no direct overlay of device trajectory versus analytic LIF solution are reported. Without these, it is impossible to judge how closely the chosen aspect ratio actually reproduces the target dynamics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the presentation of the LLG results.

read point-by-point responses

  1. Referee: [Abstract / LLG simulations] Abstract and LLG simulation section: the central claim that “proper free-layer dimensions allow the device response to follow a signed LIF equation” rests entirely on deterministic LLG runs. No Langevin thermal-noise term is included, yet at the stated 10 nm scale the kT fluctuation strength is comparable to anisotropy energy and can destroy both the linear leak regime and the bipolar threshold crossing required for the signed-LIF mapping. This omission is load-bearing for the device-to-neuron equivalence.

    Authors: We agree that thermal fluctuations represent an important consideration at these dimensions and that their absence limits claims about robustness. The deterministic LLG runs were intended to establish the existence of a dimension regime in which the orthogonal-easy-axis MTJ can produce signed-LIF-like dynamics. In the revision we will add a dedicated subsection discussing thermal effects and will include preliminary stochastic LLG simulations that incorporate the Langevin term to test whether the linear-leak and bipolar-threshold regimes survive realistic noise levels. revision: yes

  2. Referee: [Abstract] Abstract: no simulation parameters (damping, saturation magnetization, temperature, integration timestep), no goodness-of-fit metric (e.g., RMS error or R² to the ideal signed-LIF ODE), and no direct overlay of device trajectory versus analytic LIF solution are reported. Without these, it is impossible to judge how closely the chosen aspect ratio actually reproduces the target dynamics.

    Authors: We accept that the current manuscript lacks the quantitative details needed for independent assessment. The revised version will report the complete set of LLG parameters, supply RMS-error and R² values quantifying agreement with the analytic signed-LIF ODE, and add figures that overlay representative device trajectories against the ideal solution for the 10 nm × 45 nm × 50 nm geometry. revision: yes

Circularity Check

0 steps flagged

No circularity: simulation-to-model chain is independent

full rationale

The paper derives the signed-LIF mapping from deterministic LLG simulations of the orthogonal-easy-axis MTJ for chosen dimensions (10 nm × 45 nm × 50 nm), then fits a device-neuron model to those simulation traces and evaluates it on CIFAR-10/CIFAR10-DVS. No equation or claim reduces to its own inputs by construction; the LLG runs supply external dynamical evidence, the fit is a post-processing step, and the network accuracies are genuine evaluations rather than tautological predictions. No self-citations, uniqueness theorems, or ansatzes are invoked in the provided text to close the loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the LLG equation governing dynamics and the selection of specific dimensions to match the signed LIF model; no new entities are postulated.

free parameters (1)
  • free-layer dimensions = 10 nm x 45 nm x 50 nm
    Representative values of 10 nm x 45 nm x 50 nm are chosen so the response follows the signed LIF equation.
axioms (1)
  • standard math Landau-Lifshitz-Gilbert equation accurately describes the magnetic moment dynamics in the MTJ
    Invoked for all device simulations in the abstract.

pith-pipeline@v0.9.1-grok · 5673 in / 1298 out tokens · 15866 ms · 2026-06-28T07:26:12.876293+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages

  1. [1]

    Spintronic neural systems

    K. Roy, et al. “Spintronic neural systems.” Nat Rev Electr Eng vol. 1, pp. 714 –729, 2024. DOI: 10.1038/s44287-024-00107-9

    work page doi:10.1038/s44287-024-00107-9 2024

  2. [2]

    A crossbar array of magnetoresistive memory devices for in-memory computing

    S. Jung, et al. “A crossbar array of magnetoresistive memory devices for in-memory computing.” Nature vol. 601, pp. 211 –216, 2022. DOI: 10.1038/s41586-021- 04196-6

    work page doi:10.1038/s41586-021- 2022

  3. [3]

    Domain wall magnetic tunnel junction- based artificial synapses and neurons for all -spin neuromorphic hardware

    L. Liu, et al. “Domain wall magnetic tunnel junction- based artificial synapses and neurons for all -spin neuromorphic hardware.” Nat. Commun. vol. 15, no. 4534, 2024. DOI: doi.org/10.1038/s41467-024-48631-4

    work page doi:10.1038/s41467-024-48631-4 2024

  4. [4]

    Nanoscale exchange -bias magnetic tunnel junctions enabled memristive synapse and leaky - integrate-fire neuron for neuromorphic computing

    Z. Chen, et al. “Nanoscale exchange -bias magnetic tunnel junctions enabled memristive synapse and leaky - integrate-fire neuron for neuromorphic computing.” Nat. Commun. Mar. 24, 2026, to be published. DOI: 10.1038/s41467-026-70802-8. 50

    work page doi:10.1038/s41467-026-70802-8 2026

  5. [5]

    Neuromorphic Hebbian learning with magnetic tunnel junction synapses

    P. Zhou, et al. “Neuromorphic Hebbian learning with magnetic tunnel junction synapses.” Commun. Eng. vol. 4, pp. 142, 2025. DOI: 10.1038/s44172-025-00479-2

    work page doi:10.1038/s44172-025-00479-2 2025

  6. [6]

    Spiking dynamics in dual free layer perpendicular magnetic tunnel junctions,

    L. Farcis, et al. “Spiking dynamics in dual free layer perpendicular magnetic tunnel junctions,” Nano Letters, vol. 23, no. 17, pp. 7869 –7875, 2023, DOI: 10.1021/acs.nanolett.3c01597

    work page doi:10.1021/acs.nanolett.3c01597 2023

  7. [7]

    Domain Wall Leaky Integrate- and-Fire Neurons with Shape -Based Configurable Activation Functions,

    W. H. Brigner, J. Alsaleem, K. Zahedinejad, J. A. C. Incorvia, and S. Rakheja, “Domain Wall Leaky Integrate- and-Fire Neurons with Shape -Based Configurable Activation Functions,” IEEE T. Electron Dev., vol. 69, no. 5, pp. 2353 –2359, May 2022. doi: 10.1109/TED.2022.3159508. [Online]. Available: https://doi.org/10.1109/TED.2022.3159508 61

    work page doi:10.1109/ted.2022.3159508 2022

  8. [8]

    Shape Anisotropy - Dependent Leaking in Magnetic Neurons for Bio - Mimetic Neuromorphic Computing,

    T. Leonard, N. Zogbi, S. Liu, W. S. Rogers, C. H. Bennett, and J. A. C. Incorvia, “Shape Anisotropy - Dependent Leaking in Magnetic Neurons for Bio - Mimetic Neuromorphic Computing,” ACS Nano, vol. 19, no. 3, pp. 3470 –3477, 2025. doi: 10.1021/acsnano.4c13020. [Online]. Available: https://doi.org/10.1021/acsnano.4c13020

    work page doi:10.1021/acsnano.4c13020 2025

  9. [9]

    BackEISNN: A Deep Spiking Neural Network With Adaptive Self -Feedback and Balanced Excitatory –Inhibitory Neurons,

    D. Zhao, Y . Zeng, and Y . Li, “BackEISNN: A Deep Spiking Neural Network With Adaptive Self -Feedback and Balanced Excitatory –Inhibitory Neurons,” Neural Networks, vol. 154, pp. 68 –77, 2022. doi: 10.1016/j.neunet.2022.06.036. [Online]. Available: https://doi.org/10.1016/j.neunet.2022.06.036

    work page doi:10.1016/j.neunet.2022.06.036 2022

  10. [10]

    An All Integer -Based Spiking Neural Network With Dynamic Threshold Adaptation,

    C. Zou, X. Cui, S. Feng, G. Chen, Y . Zhong, Z. Dai, and Y . Wang, “An All Integer -Based Spiking Neural Network With Dynamic Threshold Adaptation,” Frontiers in Neuroscience, vol. 18, Art. no. 1449020, 2024. doi: 10.3389/fnins.2024.1449020. [Online]. Av ailable: https://doi.org/10.3389/fnins.2024.1449020 76

    work page doi:10.3389/fnins.2024.1449020 2024

  11. [11]

    Fast-snn: Fast spiking neural network by converting quantized ann

    Y . Hu, et al. “Fast-snn: Fast spiking neural network by converting quantized ann.” IEEE Trans. Pattern Anal. Mach. Intell. vol. 45, no. 12, pp. 14546-14562, Dec. 2023, DOI: 10.1109/TPAMI.2023.3275769

    work page doi:10.1109/tpami.2023.3275769 2023

  12. [12]

    Size dependence of intrinsic spin transfer switching current density in elliptical spin valves

    R. Heindl, et al. “Size dependence of intrinsic spin transfer switching current density in elliptical spin valves.” Appl. Phys. Lett. vol. 92, no. 26, 2008. DOI: 10.1063/1.2953980

    work page doi:10.1063/1.2953980 2008

  13. [13]

    Magnetization Dynamics Modulated by Dzyaloshinskii-Moriya Interaction in the Double - Interface Spin -Transfer Torque Magnetic Tunnel Junction

    S. Li, et al. “Magnetization Dynamics Modulated by Dzyaloshinskii-Moriya Interaction in the Double - Interface Spin -Transfer Torque Magnetic Tunnel Junction.” Nanoscale Research Letters vol. 14, no. 315,

  14. [14]

    DOI: 10.1186/s11671-019-3150-4. 86

    work page doi:10.1186/s11671-019-3150-4

  15. [15]

    Magnetic Tunnel Junction Mimics Stochastic Cortical Spiking Neurons

    A. Sengupta, et al. “Magnetic Tunnel Junction Mimics Stochastic Cortical Spiking Neurons.” Scientific Reports vol. 6, no. 30039, 2016. DOI: 10.1038/srep30039

    work page doi:10.1038/srep30039 2016

  16. [16]

    A Compact Model of Perpendicular Spin -Transfer-Torque Magnetic Tunnel Junction,

    C. -T. Tung, et al. “A Compact Model of Perpendicular Spin -Transfer-Torque Magnetic Tunnel Junction,” IEEE T. Electron Dev., vol. 71, no. 1, pp. 57– 61, Jan. 2024, doi: 10.1109/TED.2023.3313997

    work page doi:10.1109/ted.2023.3313997 2024

  17. [17]

    Spin Transfer Torque Switching Dynamics in CoFeB/MgO Magnetic Tunnel Junctions

    A. Meo, et al. “Spin Transfer Torque Switching Dynamics in CoFeB/MgO Magnetic Tunnel Junctions.” Physical Review B, vol. 103, no. 5, Art. no. 054426, 2021. DOI: 10.1103/PhysRevB.103.054426

    work page doi:10.1103/physrevb.103.054426 2021

  18. [18]

    Deep Residual Learning for Image Recognition

    K. He, et al. “Deep Residual Learning for Image Recognition.” Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770 –778,

  19. [19]

    DOI: 10.1109/CVPR.2016.90

    work page doi:10.1109/cvpr.2016.90 2016

  20. [20]

    Spikformer: When Spiking Neural Network Meets Transformer

    Z. Zhou, et al. “Spikformer: When Spiking Neural Network Meets Transformer.” Proc. International Conference on Learning Representations (ICLR), 2023

    2023