Neuron Polarization Effect Overview

• Neuron Polarization Effect is an umbrella term for diverse polarity-sensitive phenomena ranging from extracellular field modulation in neurons to scale-dependent selectivity in artificial neural networks.
• Studies reveal that neuronal geometry, membrane asymmetry, and tissue heterogeneity can amplify extracellular voltages—up to 60% near the axon initial segment—modulating excitability.
• The effect is extended to device analogues and optical models, highlighting its role in reproducing depolarization/repolarization dynamics and population divergence in computational frameworks.

Searching arXiv for the cited works to ground the article in published preprints. arXiv search: "Neuron Polarization Effect" In the available arXiv literature, the expression “Neuron Polarization Effect” does not denote a single universally standardized phenomenon. Depending on the domain, it refers to changes in neuronal membrane polarization and extracellular field structure, polarity-dependent shifts in excitability under electrical stimulation, electromechanical asymmetry of polar membranes, polarization-preserving or polarization-degrading optical propagation in myelinated axons, slowly propagating cortical dipole-polarization waves, and, in large artificial neural networks, a scale-dependent separation between more selective and less selective neuron populations (Bauer et al., 2012, Mosgaard et al., 2014, Davies et al., 7 May 2026, Dravid et al., 2 Jun 2026). The common thread is a transition from a symmetric or weakly structured state to a state in which polarity, selectivity, or directional asymmetry becomes dynamically consequential; the state variable itself, however, differs sharply across literatures.

1. Terminology and domains of use

The term spans several technically distinct research traditions. In biophysical neuroscience, polarization usually concerns membrane voltage, extracellular potential gradients, or dipole polarization density. In membrane physics, it denotes spontaneous electric polarization and its offset potential. In optical models, it refers to the polarization state of guided light. In machine learning, it is a population-level interpretability effect in artificial neurons rather than an electrophysiological variable.

Domain	Polarized quantity	Representative papers
Extracellular and membrane electrophysiology	Extracellular potential, membrane polarization, excitability	(Bauer et al., 2012, Wang et al., 2018, Mosgaard et al., 2014)
Theoretical neurophysiology and psychology	Stationary depolarization, repolarization, hyperpolarization as indices of functional-metabolic state	(Murik, 2011)
Neuromorphic and iontronic devices	Ion-modulated conductance states, antiambipolar transport, induced electric fields	(Harikesh et al., 2022, Santos et al., 2023, Wu et al., 2019)
Optical and cortical-wave formulations	Optical polarization fidelity, cortical dipole-polarization waves	(Davies et al., 7 May 2026, Jang et al., 7 Mar 2026)
Artificial neural networks	Population-level selectivity and monosemanticity divergence	(Dravid et al., 2 Jun 2026)

This distribution of meanings suggests that the phrase is intrinsically context-dependent. In technical usage, the relevant state variable must therefore be specified explicitly.

2. Extracellular field generation and polarization-relevant neuronal coupling

A central biophysical formulation treats neuron polarization effects as consequences of the extracellular fields generated by transmembrane currents. In a finite-element study of CA1 pyramidal neurons adapted from Pinsky and Rinzel and augmented with $I_h$ and $I_m$, compartmental membrane currents were exported into COMSOL as boundary current sources $Q_j(t)$, and the extracellular potential $V$ was computed from a Maxwell-derived quasistatic potential equation,

$E=-\nabla V,$

$$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$

The key question was how neuronal geometry, orientation, and heterogeneous extracellular resistivity alter extracellular potentials relative to the standard homogeneous-medium approximation. The study emphasized hippocampal laminar heterogeneity, especially the experimental result that the stratum pyramidale is about twice as resistive as surrounding layers according to Lopez-Aguado et al., and showed that including this heterogeneity substantially increases local field amplitudes near emitting neurons (Bauer et al., 2012).

The quantitative effect was strongest near the axon initial segment. For a single neuron during the peak of an action potential of about $80\ \mathrm{mV}$, the largest extracellular voltage in the homogeneous case occurred near $AIS_2$, with peak $0.25\ \mathrm{mV}$. When heterogeneous resistivity was included, values measured in point probes within stratum pyramidale became 60% larger at close distance than in the homogeneous case and 28% higher as a spatial mean over the full $1\ \mu\mathrm{m}$ to $I_m$0 distance range. In lower-resistivity layers the mean increases were smaller, about 4% in stratum oriens and 7% in stratum radiatum. In multi-neuron arrangements, four neighboring active neurons at $I_m$1 interaxonal spacing produced extracellular voltages of $I_m$2 for asynchronous firing and $I_m$3 for synchronous firing in heterogeneous tissue, compared with $I_m$4 and $I_m$5 in the homogeneous case. In an oblique geometry with nearest interaxonal distance $I_m$6 near $I_m$7, heterogeneous-medium values reached $I_m$8 for asynchronous activity and $I_m$9 for synchronous activity. The study therefore supports the claim that source synchrony, crowding, orientation, and laminar resistivity all shape polarization-relevant extracellular voltages.

The same work also states an important limitation: it does not directly simulate a second active neuron’s membrane dynamics in a bidirectionally coupled way. It computes extracellular fields from one or more active neurons and evaluates the resulting extracellular voltages at points or at nearby passive neuronal structures. Its strongest conclusion is therefore about field amplitude estimation, not a direct demonstration of ephaptic spike generation. The authors note that their FEM simulations did not produce spiking solely from extracellular fields generated by neighboring neurons, supporting modulatory rather than spike-triggering effects under the simulated conditions.

3. Membrane asymmetry, stimulation polarity, and functional-state interpretations

A second line of work frames the neuron polarization effect at the membrane level. In a general theory of polar biological membranes, the membrane is treated as a capacitor,

$Q_j(t)$0

whose charge and free energy become asymmetric when the membrane has a spontaneous polarization $Q_j(t)$1. For symmetric membranes, the electrical free energy is quadratic in voltage and positive and negative voltages of the same magnitude have identical electrostrictive effects. For asymmetric membranes, total polarization is

$Q_j(t)$2

and the membrane charge and electrical free energy become

$Q_j(t)$3

$Q_j(t)$4

The offset potential $Q_j(t)$5 means the capacitor is discharged at a nonzero applied voltage, not at $Q_j(t)$6. This yields unequal responses to depolarization and hyperpolarization, voltage-dependent electrostriction, and, in the paper’s formulation, outward or inward rectification of membrane permeability (Mosgaard et al., 2014).

A distinct but related stimulation literature treats polarization operationally as a polarity-dependent shift in effective excitability. In an acute rat Tibialis Anterior preparation, enveloped high frequency stimulation (EHFS) was designed so that positive and negative electrodes recruit neurons more synchronously than with conventional square pulses. The paper’s mechanistic claim is that positive external stimulation facilitates larger $Q_j(t)$7 influx and smaller $Q_j(t)$8 efflux, thereby reducing threshold for subsequent stimulation, whereas negative external stimulation facilitates smaller $Q_j(t)$9 influx and larger $V$0 efflux, thereby increasing threshold. EHFS used high-frequency inner sinewave current pulses under a lower-frequency outer sinewave envelope; in the main experiments the inner frequency was typically 20 kHz, each envelope lasted 1 ms, one envelope was delivered every 16.7 ms, and a train of 10 envelopes was delivered every 1 s. Under continuous EHFS, force output could decrease to zero and then recover or increase, which the authors interpreted as a dynamic excitability shift produced by stimulation-induced disturbance of transmembrane ion distribution (Wang et al., 2018).

That study is also careful about its evidentiary level. It does not directly record intracellular membrane potential, extracellular $V$1, or ion-channel state variables. Its strongest evidence concerns threshold shifts, recruitment probability, and force output, and it explicitly warns that its terminology differs from classical anode–cathode rules because it focuses on the subsequent effect of sustained positive versus negative external stimulation on future excitability rather than on the instantaneous extracellular depolarization pattern.

A third, more speculative usage appears in Murik’s “polarization theory,” which interprets stationary membrane potential as an integral index of neuronal functional and metabolic state. In that framework, stationary depolarization indicates an unfavorable state with reduced lability and metabolic strain, repolarization indicates restoration, and hyperpolarization indicates a good or excellent functional state with increased lability and adaptive readiness. Negative steady-potential shifts are interpreted as depolarization, positive shifts as repolarization or hyperpolarization, and the theory uses paradigms such as ischemia, Nembutal narcosis, and hunger-driven activity to connect membrane polarization to motivation, emotion, and attention (Murik, 2011). The paper itself acknowledges that the step from slow electrical shifts to subjective experience is interpretive rather than directly proven.

4. Artificial and materials analogues of neuronal polarization

Several device-oriented papers reproduce neuron-like polarization dynamics without modeling a biological membrane directly. A conductance-based organic electrochemical neuron (c-OECN) implements a membrane-voltage analog $V$2 by combining a Na-OECT and a K-OECT built from the ladder-type mixed ion-electron conducting polymer BBL. The governing current balance is written as

$V$3

and the paper interprets the two sides of the BBL device’s Gaussian antiambipolar transfer curve as analogs of Na-channel activation and Na-channel inactivation, while the K-OECT provides delayed activation. The resulting circuit reproduces a resting $V$4 of about $V$5, a depolarized peak of about $V$6, repolarization, hyperpolarization, refractory-like behavior, and firing frequencies from about 5 Hz with $V$7 and $V$8 up to 80 Hz, with the abstract stating frequencies nearing 100 Hz. The device also shows $V$9-based modulation and GABA-induced inhibition (Harikesh et al., 2022).

The material mechanism behind that artificial polarization cycle is electrochemical doping. In BBL, conductivity first increases with doping and then decreases at high electrochemical doping, above roughly $E=-\nabla V,$0 electron/monomer, because multiply charged species with reduced mobility form. The paper treats this stable, reversible antiambipolarity as the physical origin of the Na-like activation/inactivation analog. This is a circuit-level reproduction of depolarization and repolarization, not a full ion-diffusion model of a neuronal membrane.

A more indirect materials analogue appears in a study of nanoconfined slit electrolytes between polarizable walls. There, current and conductivity are higher when the confining walls behave as conductors than when they behave as dielectrics, the current–field relation is nonlinear in both cases, and the known linear Ohmic result is recovered only in the two-dimensional limit between conductors. The paper emphasizes that the geometric location of the induced polarization charge plane matters strongly: explicit simulations compare placing the polarization charge on the material surface, $E=-\nabla V,$1, or inside the walls, $E=-\nabla V,$2, and moving the polarization plane inward damps adsorption, pairing differences, and transport modulation (Santos et al., 2023). The connection to biological neuron polarization is indirect; the paper is framed around mixed ionic-electronic conductors, memristors, and neuromorphic iontronics rather than living neurons.

A third device proposal uses an antiferromagnet/ferromagnet spintronic nanostructure as a Spin-orbit Torque Neurostimulator. In that model, switching of perpendicular magnetization in an AFM/FM bilayer generates a time-varying magnetic stray field, and Maxwell–Faraday induction produces local electric-field pulses that are proposed to modulate nearby voltage-gated sodium channels. OOMMF simulations predict induced fields of about $E=-\nabla V,$3–$E=-\nabla V,$4 for a 14 nm pillar and $E=-\nabla V,$5–$E=-\nabla V,$6 for a 112 nm pillar, with pulse widths of about 2–3 ns (Wu et al., 2019). The paper’s interpretation is explicitly excitatory and depolarizing, but it also states that no experimental neural validation, no direct membrane model, and no explicit depolarization–hyperpolarization mapping were provided.

5. Optical, cortical-wave, and reaction–diffusion extensions

Some recent work uses the language of polarization in a non-electrophysiological sense. In a biophotonic model of myelinated axons as optical waveguides, the relevant quantity is the polarization fidelity of guided light rather than membrane voltage. A $E=-\nabla V,$7 axon with four nodes of Ranvier was simulated in 3D FDTD with refractive indices $E=-\nabla V,$8, $E=-\nabla V,$9, $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$0, and wavelength $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$1. Fidelity was defined as the normalized overlap between the input electric field profile and the field at position $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$2,

$$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$3

In that framework, variation in myelin thickness alone had minimal impact, non-circular cross-sections showed strong mode dependence, and axonal bending had the most significant influence, producing large fluctuations and deep fidelity dips. When all imperfections were combined, certain modes still exhibited repeated revivals to around $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$4, stronger than the revivals in the single-imperfection cases (Davies et al., 7 May 2026). This is an optical polarization effect in neurons, not an electrophysiological one.

A separate cortical-wave formulation treats polarization as dipole polarization density in tissue. For primary visual cortex, impressed ionic currents and Debye-type dipole relaxation lead to a telegraph-type scalar-potential equation,

$$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$5

The paper then argues that the scalar potential field $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$6 and a polarization wave $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$7 are related by linear convolution,

$$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$8

so both propagate with the same velocity. Using a modulation-based estimate and ridge tracking in simulations, the reported propagation speed is about $$-\nabla \cdot \left(\sigma \nabla V + \varepsilon_0 \varepsilon_r \frac{\partial}{\partial t}\nabla V \right)=Q_j.$$9, close to cited cortical travelling-wave speeds. The same work further argues that multi-$80\ \mathrm{mV}$0 polarization waves undergo dispersive spreading in time, which possibly suppresses cross-channel interference in visual perception (Jang et al., 7 Mar 2026). Here again, “polarization” refers to slowly oscillating neuronal dipoles and cortical scalar-potential propagation rather than to membrane depolarization.

A broader polarity literature concerns reaction–diffusion patterning rather than electrical polarization. In a one-dimensional antagonistic network model of cell polarization, a polarized pattern collapses into a homogeneous state when subjected to single-sided self-regulation, single-sided additional regulation, or unequal system parameters. Stable polarity can be restored by combining two modifications with opposing effects, and spatially inhomogeneous parameters can pin the interface at a designated location (Chen et al., 2024). That paper is about cell polarity in general rather than neurons. A plausible implication is that stable neuronal asymmetry may likewise require a balance of opposing feedbacks rather than unopposed local amplification.

6. Representation-level polarization in artificial neural networks

In large artificial neural networks, the phrase has acquired a fully different technical meaning. The paper “Neuron Populations Exhibit Divergent Selectivity with Scale” defines the Neuron Polarization Effect as a scale-dependent separation between Rosetta Neurons, whose activation patterns recur across independently trained models, and a complementary non-Rosetta population. The empirical claim is that scaling does not make all neurons uniformly more interpretable; instead, a shared Rosetta population becomes increasingly selective, monosemantic, and specialized, while the remaining neurons form a larger, less selective, more polysemantic background (Dravid et al., 2 Jun 2026).

The analysis is restricted to MLP neurons in Transformer blocks. Rosetta Neurons are identified by Pearson-correlation similarity of aligned activations across model pairs and mutual nearest-neighbor matching. Their count follows a sublinear power law,

$80\ \mathrm{mV}$1

with fitted exponents in both language and vision of approximately $80\ \mathrm{mV}$2–$80\ \mathrm{mV}$3 and $80\ \mathrm{mV}$4 around $80\ \mathrm{mV}$5. Because the exponent is sublinear, Rosetta Neurons grow in absolute number but occupy a shrinking fraction of total neurons. Under a dataset-permutation null, Rosetta counts collapse to roughly 20–100 matches and lose systematic scaling. In LLMs, selectivity is measured by excess kurtosis of vocabulary-space output projections,

$80\ \mathrm{mV}$6

and Rosetta Neurons show increasing mean excess kurtosis with scale while non-Rosetta neurons remain near zero. In vision models, a VLM-as-a-judge procedure finds that the fraction of Rosetta Neurons judged monosemantic increases with scale, while the corresponding non-Rosetta fraction decreases.

The analytical model in that paper formalizes polarization as a capacity-allocation effect under superposition. Feature importance is assumed to follow

$80\ \mathrm{mV}$7

and neuron isolation scores are allocated under a linear budget,

$80\ \mathrm{mV}$8

The resulting Rosetta count scales as

$80\ \mathrm{mV}$9

while average isolation increases in the Rosetta population and tends to zero in the non-Rosetta tail. This usage is conceptually distinct from biological neuron polarization, but it preserves the same higher-level notion of divergence between two populations under scale.

The collected literature therefore supports a strongly plural definition of the neuron polarization effect. In biological electrophysiology it concerns membrane voltage, extracellular field structure, and polarity-dependent excitability; in membrane physics it concerns permanent polarization and offset potentials; in device research it concerns ion-modulated conductance trajectories that emulate depolarization and repolarization; in optical and cortical-wave studies it concerns polarization of light or dipoles; and in machine learning it names a scaling law of selective versus superposed artificial neurons. This suggests that the phrase is best treated not as a single phenomenon, but as a family of polarity-sensitive effects whose meaning depends on the governing variable, the mathematical formalism, and the measurement regime.