pith. sign in

arxiv: 2604.03187 · v1 · submitted 2026-04-03 · 💻 cs.NE · cond-mat.other· physics.app-ph

Biologically Realistic Dynamics for Nonlinear Classification in CMOS+X Neurons

Steven Louis , Hannah Bradley , Artem Litvinenko , Cody Trevillian , Darrin Hanna , Vasyl Tyberkevych This is my paper

Pith reviewed 2026-05-13 18:29 UTC · model grok-4.3

classification 💻 cs.NE cond-mat.otherphysics.app-ph
keywords spiking neural networksmagnetic tunnel junctionnonlinear computationXOR classificationneuromorphic hardwareCMOS+X neuronthreshold activationresponse latency
0
0 comments X

The pith

Circuit simulations show three intrinsic properties of MTJ-based CMOS+X neurons enable nonlinear XOR classification without added circuits.

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

This paper examines nonlinear computation in a compact spiking neuron made from a magnetic tunnel junction connected in series with an NMOS transistor. Simulations of a multilayer network solving the XOR problem demonstrate that threshold activation, response latency, and absolute refraction together produce the required nonlinearity. Threshold activation selects participating neurons, response latency shifts spike timing, and absolute refraction prevents repeated spikes. A sympathetic reader would care because spiking networks target energy-efficient artificial intelligence, and these properties arise naturally from the device dynamics rather than from extra circuitry that would increase power use.

Core claim

In circuit simulations of a multilayer spiking network solving the XOR classification problem, the CMOS+X neuron exhibits nonlinear behavior through three intrinsic properties arising from MTJ magnetization dynamics: threshold activation that determines which neurons participate in computation, response latency that shifts spike timing, and absolute refraction that suppresses subsequent spikes. These properties together allow the compact hardware to achieve expressive computation without relying on additional circuit complexity.

What carries the argument

The three intrinsic neuronal properties—threshold activation, response latency, and absolute refraction—produced by the magnetization dynamics of the magnetic tunnel junction in series with an NMOS transistor.

If this is right

  • Nonlinear tasks become feasible in compact neuromorphic hardware that avoids extra circuitry for energy efficiency.
  • Multilayer spiking networks can leverage MTJ device dynamics for classification problems such as XOR.
  • Threshold activation, latency shifts, and refraction suppression together replace the need for explicit nonlinear activation functions.
  • Magnetization dynamics in MTJs can directly support expressive computation in spiking systems.

Where Pith is reading between the lines

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

  • If the properties translate from simulation to fabricated devices, power consumption in edge AI hardware could drop by eliminating auxiliary nonlinear circuits.
  • The same three properties might generalize to other nonlinearly separable tasks beyond XOR, enabling broader use in compact spiking networks.
  • Fabrication experiments could reveal whether real MTJ noise or variability preserves or disrupts the timing-based nonlinearity seen in simulation.

Load-bearing premise

The circuit simulations accurately capture real MTJ magnetization dynamics and that these three properties alone suffice for the observed nonlinear classification without unmodeled effects.

What would settle it

Fabricating a physical CMOS+X neuron circuit and testing whether it solves the XOR problem with nonlinear accuracy matching the simulations, or whether it reverts to linear behavior under varied input conditions.

Figures

Figures reproduced from arXiv: 2604.03187 by Artem Litvinenko, Cody Trevillian, Darrin Hanna, Hannah Bradley, Steven Louis, Vasyl Tyberkevych.

Figure 1
Figure 1. Figure 1: Electrical schematic illustrating two NMOS+MTJ neurons connected [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Simulation results of rows 1 and 3 of the XOR truth table. (a,e) Input [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation result of row 2 of the XOR truth table. (a) Input currents [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
read the original abstract

Spiking neural networks encode information in spike timing and offer a pathway toward energy efficient artificial intelligence. However, a key challenge in spiking neural networks is realizing nonlinear and expressive computation in compact, energy-efficient hardware without relying on additional circuit complexity. In this work, we examine nonlinear computation in a CMOS+X spiking neuron implemented with a magnetic tunnel junction connected in series with an NMOS transistor. Circuit simulations of a multilayer network solving the XOR classification problem show that three intrinsic neuronal properties enable nonlinear behavior: threshold activation, response latency, and absolute refraction. Threshold activation determines which neurons participate in computation, response latency shifts spike timing, and absolute refraction suppresses subsequent spikes. These results show that magnetization dynamics of MTJ devices can support nonlinear computation in compact neuromorphic hardware.

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

1 major / 2 minor

Summary. The paper claims that circuit simulations of a multilayer spiking network using CMOS+X neurons (MTJ in series with NMOS transistor) solve the XOR classification task, and that this nonlinear behavior arises specifically from three intrinsic properties of the device physics: threshold activation, response latency, and absolute refraction. These properties are said to determine neuron participation, shift spike timing, and suppress subsequent spikes, respectively, enabling compact neuromorphic computation without added circuit complexity.

Significance. If the attribution to these three properties holds under rigorous isolation, the result would be significant for energy-efficient neuromorphic hardware, as it shows how MTJ magnetization dynamics can supply the nonlinearity required for expressive spiking networks in a compact, low-power form factor.

major comments (1)
  1. [XOR network simulation results] The central claim that threshold activation, response latency, and absolute refraction are the causal drivers of nonlinear XOR classification is not supported by ablation or control simulations. No experiments are shown that disable one property at a time (e.g., zero refractory period or fixed latency) while retaining the identical MTJ-CMOS circuit and Landau-Lifshitz-Gilbert dynamics; without such checks, other aspects of the full magnetization or transistor model could be responsible for the observed timing-based nonlinearity.
minor comments (2)
  1. [Abstract and results] The abstract and results sections report no error bars, quantitative classification accuracy, or parameter sensitivity analysis on the circuit simulations.
  2. [Methods] No comparison is provided between the simulated MTJ switching behavior and measurements from fabricated devices to validate the model assumptions.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the causal attribution to the three properties would be strengthened by explicit ablation studies and will incorporate such controls in the revised version.

read point-by-point responses

  1. Referee: [XOR network simulation results] The central claim that threshold activation, response latency, and absolute refraction are the causal drivers of nonlinear XOR classification is not supported by ablation or control simulations. No experiments are shown that disable one property at a time (e.g., zero refractory period or fixed latency) while retaining the identical MTJ-CMOS circuit and Landau-Lifshitz-Gilbert dynamics; without such checks, other aspects of the full magnetization or transistor model could be responsible for the observed timing-based nonlinearity.

    Authors: We acknowledge that the original manuscript lacks explicit ablation or control simulations that isolate each property while preserving the core MTJ-CMOS circuit and LLG dynamics. To address this rigorously, we will add new simulation results in the revision. Specifically, we will (1) disable absolute refraction by setting the post-spike refractory period to zero in the model while retaining all other LLG and transistor parameters; (2) eliminate response latency by removing the integration time constant to produce instantaneous spiking; and (3) remove threshold activation by replacing the nonlinear threshold with a linear response function. These controls will be performed on the identical circuit topology and magnetization dynamics. The results will show degraded XOR performance when any single property is removed, thereby confirming their causal roles in the observed nonlinearity. revision: yes

Circularity Check

0 steps flagged

No circularity: simulation results derive from independent physical models

full rationale

The paper reports circuit simulations of MTJ-CMOS neurons performing XOR classification and attributes the nonlinearity to three emergent properties (threshold activation, response latency, absolute refraction) arising from standard Landau-Lifshitz-Gilbert magnetization dynamics coupled to NMOS transistor equations. These properties are not defined in terms of the classification task, nor are any parameters fitted to the XOR outcome and then presented as predictions. No self-citations supply load-bearing uniqueness theorems or ansatzes; the models are externally standard and falsifiable against known MTJ switching data. The derivation chain therefore remains self-contained against independent device physics benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim depends on standard domain assumptions about MTJ magnetization dynamics and NMOS transistor models; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption MTJ devices obey standard Landau-Lifshitz-Gilbert magnetization dynamics coupled to spin-transfer torque
    Invoked implicitly to justify the threshold, latency, and refraction behaviors in the circuit simulations.

pith-pipeline@v0.9.0 · 5441 in / 1166 out tokens · 26790 ms · 2026-05-13T18:29:29.587802+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    An overview on large language models across key domains: a systematic review,

    M. Bruscia, G. A. Manduzio, F. A. Galatolo, M. G. Cimino, A. Greco, L. Cominelli, and E. P. Scilingo, “An overview on large language models across key domains: a systematic review,” in2024 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE). IEEE, 2024, pp. 125–130

    work page 2024

  2. [2]

    Cmos-compatible neuromorphic devices for neuromorphic perception and computing: a review,

    Y . Zhu, H. Mao, Y . Zhu, X. Wang, C. Fu, S. Ke, C. Wan, and Q. Wan, “Cmos-compatible neuromorphic devices for neuromorphic perception and computing: a review,”International Journal of Extreme Manufacturing, vol. 5, no. 4, p. 042010, 2023

    work page 2023

  3. [3]

    In-datacenter performance analysis of a tensor processing unit,

    N. P. Jouppi, C. Young, N. Patil, D. Patterson, G. Agrawal, R. Bajwa, S. Bates, S. Bhatia, N. Boden, A. Borcherset al., “In-datacenter performance analysis of a tensor processing unit,” inProceedings of the 44th annual international symposium on computer architecture, 2017, pp. 1–12

    work page 2017

  4. [4]

    Networks of spiking neurons: the third generation of neural network models,

    W. Maass, “Networks of spiking neurons: the third generation of neural network models,”Neural networks, vol. 10, no. 9, pp. 1659–1671, 1997

    work page 1997

  5. [5]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jainet al., “Loihi: A neuromorphic manycore processor with on-chip learning,”Ieee micro, vol. 38, no. 1, pp. 82–99, 2018

    work page 2018

  6. [6]

    A cmos+ x spiking neuron with on-chip machine learning,

    S. Louis, M. B. Abramson, H. Bradley, C. Trevillian, G. D. Nelson, A. Slavin, A. Litvinenko, J. Gorski, I. N. Krivorotov, D. Hannaet al., “A cmos+ x spiking neuron with on-chip machine learning,”arXiv preprint arXiv:2512.03966, 2025

    work page arXiv 2025

  7. [7]

    Squire, D

    L. Squire, D. Berg, F. E. Bloom, S. Du Lac, A. Ghosh, and N. C. Spitzer, Fundamental neuroscience. Academic press, 2012

    work page 2012

  8. [8]

    Gerstner and W

    W. Gerstner and W. M. Kistler,Spiking neuron models: Single neurons, populations, plasticity. Cambridge university press, 2002

    work page 2002

  9. [9]

    A 0.086-mm 2 12.7- pj/sop 64k-synapse 256-neuron online-learning digital spiking neuro- morphic processor in 28-nm cmos,

    C. Frenkel, M. Lefebvre, J.-D. Legat, and D. Bol, “A 0.086-mm 2 12.7- pj/sop 64k-synapse 256-neuron online-learning digital spiking neuro- morphic processor in 28-nm cmos,”IEEE Transactions on Biomedical Circuits and Systems, vol. 13, no. 1, pp. 145–158, 2019

    work page 2019

  10. [10]

    Spintronic neuron using a magnetic tunnel junction for low-power neuromorphic computing,

    S. Louis, H. Bradley, C. Trevillian, A. Slavin, and V . Tyberkevych, “Spintronic neuron using a magnetic tunnel junction for low-power neuromorphic computing,”IEEE Magnetics Letters, vol. 15, pp. 1–5, 2024

    work page 2024

  11. [11]

    A physics- based circuit model for magnetic tunnel junctions,

    S. Louis, H. Bradley, A. Litvinenko, and V . Tyberkevych, “A physics- based circuit model for magnetic tunnel junctions,”IEEE Magnetics Letters, 2025

    work page 2025

  12. [12]

    Everspin — The MRAM Company — everspin.com,

    “Everspin — The MRAM Company — everspin.com,” https://www.everspin.com/, [Accessed 01-03-2026]

    work page 2026