Intrinsic noise is a neuromodulator with two dials, and one of them is a cliff
Engineers treat noise as the enemy of computation; biologists know neurons are stochastic down to the ion channel and use it anyway. A new analysis of a memristive neuron model makes that trade-off precise: noise in the fast excitatory feedback creates, tunes, and beyond a critical level destroys excitability, while noise in the slow inhibitory feedback becomes an internal stimulus that drives stochastic spiking.
Source: Neuromodulation Through Intrinsic Noise, Koo, Lee, and Sepulchre, arXiv:2611.13780, 2026-11-10. Primary source. Read the full arXiv HTML including propositions, proofs, and simulation parameters.
What the work claims
This is a theory paper: four propositions about a simplified memristive neuron, backed by graphical analysis and numerical simulation, with no new experimental data1. The model, from Burger, Shahhosseini, and Sepulchre (CDC 2025)2, is a single compartment with capacitance C, a leak conductance, an excitatory and an inhibitory current whose conductances are memductances with sigmoidal activation functions, and two internal states: a fast excitatory variable (timescale εe = 0.03 ms in the simulations) and a slow inhibitory variable (εi = 30 ms). Intrinsic noise - the random gating of ion channels, abstracted - is added independently to each pathway, with intensities σe and σi.
The claims decompose cleanly by pathway. Noise in the fast excitatory feedback smooths the nonlinear current-voltage profile that gives the neuron its memory; weak noise sharpens sensitivity, intermediate noise creates a qualitatively new behavior called anode break excitability, and strong noise destroys excitability outright at a critical intensity. Noise in the slow inhibitory feedback does none of that; instead it directly kicks the neuron across threshold, but only when anode break excitability is present. Together the two noise sources produce stochastic resonance - enhanced detection of a subthreshold periodic signal at intermediate noise - through distinct and complementary mechanisms.
How it works
Excitability in this model rests on a reverse-N-shaped fast current profile: the fold in that curve is what lets a small push trigger a large excursion, i.e. a spike. The first proposition formalizes what noise does to the fold. Under a mild conductance condition (gl/ḡe greater than about 0.1289), there exists a critical σe* at which the smoothed profile undergoes a hysteresis singularity: below it the system keeps its two threshold crossings, at it they merge, above it they vanish, and with them the memory and the excitability. Noise, in other words, is not a perturbation around excitability; past a calculable point it abolishes the category.
Before that cliff, intermediate σe buys something: anode break excitability (ABE), the tendency to fire when an inhibitory current that was holding the cell hyperpolarized is suddenly removed. The second proposition gives a crisp graphical criterion - the neuron is anode break excitable precisely when its noise-smoothed fast current intersects the voltage axis only once. The slow inhibitory state, having been dragged below rest, acts as a stored deflection; release lets the fast dynamics carry the membrane through threshold. The authors also show the effect is not model-specific: a Hodgkin-Huxley neuron with one parameter set exhibits ABE and shows stochastic resonance to a 2 sin(2πt/50) µA/cm2 subthreshold input, while shifting only the leak equilibrium potential (Vl = 2.72 mV) abolishes it.
The two dials then diverge in function. E-noise reshapes the fast profile, which continuously shifts the resting potential, the spike threshold, and the ABE threshold - a sensitivity control, the neuron's propensity to spike. I-noise, in the slow inhibitory pathway, fluctuates the internal state that ABE reads out, so it acts as an endogenous stimulus generator; proposition four proves stochastic spiking occurs under sufficient timescale separation if and only if the ABE criterion holds. Numerically, the spontaneous spiking rate rises with both intensities, and the SNR of a subthreshold 12.5 Hz signal peaks at intermediate values of both - textbook stochastic resonance, but now with a division of labor: e-noise sets how sensitively the cell listens, i-noise sets how often it spontaneously tries. In the high-SNR region (σe = 0.4, σi = 0.6) spikes synchronize with the subthreshold input, and the firing rate matches the signal frequency, consistent with the rate-matching account of resonance.
Where a skeptic should push
The load-bearing assumption is timescale separation: the propositions are proven for the reduced system in which the fast and slow variables are assumed well separated, and the dimensionless simulation parameters (εe = 0.03, εi = 30) sit comfortably in that regime. Real cortical neurons, with dendritic and axonal compartments and perisomatic inhibition arriving on multiple timescales, are not obviously one-fast-one-slow. The authors flag network extension as future work, which is where the interesting failure modes live: in a network, one neuron's i-noise is another neuron's synaptic input.
Second, the quantitative resonance result is a single numerical sweep over one parameter set. "SNR maximized at intermediate noise" is the expected shape, and the two-dial mechanism is the paper's real contribution, but the specific optimal (0.4, 0.6) should not be quoted as a biological constant; the model is dimensionless, and mapping σ onto channel-count noise in tissue is not attempted.
Third, the independent-dials framing is a modeling convenience. In tissue you cannot dial σe and σi separately; pharmacology, channel expression, and temperature move both, and disease states move both in correlated ways. The paper proves the knobs exist and are separable in principle. Whether any biological or neuromorphic system gets to turn them independently is an empirical question the paper deliberately does not answer.
Why organoid noise is a dial, not a bug
Biological computing substrates come with noise you cannot design away: ion channel stochasticity in a few-hundred-unit organoid network dwarfs the thermal noise budgets of silicon. The usual response is to average it out - long integration windows, population readouts, error-correcting codes. This paper sketches the opposite playbook, and it has three concrete entries.
First, the sensitivity dial suggests closed-loop stimulation should be tuned against the substrate's intrinsic noise level, not against an idealized deterministic neuron. Stochastic resonance means a subthreshold patterned drive - below the threshold a deterministic model would say is useless - can entrain spikes when the noise sits in the intermediate band, and the resonant firing rate locks to the signal frequency. For an organoid interface, the practical protocol is a noise-calibration sweep: estimate the spontaneous spiking statistics, then place the stimulation amplitude at the resonance peak rather than above threshold. Cheaper stimulation, less charge injection, and the tissue does part of the detection work.
Second, anode break excitability is an underused stimulation primitive. If a hyperpolarizing pulse followed by release can trigger a spike - the stored deflection in the slow inhibitory state doing the work - then brief inhibitory-then-release waveforms are a spike-triggering protocol that spends less depolarizing charge than a conventional biphasic excitatory pulse. Less charge per spike means less electrode polarization, less tissue stress, and longer-lived cultures in closed-loop rigs. The paper's own Hodgkin-Huxley demonstration, where only the leak potential distinguishes resonant from silent regimes, is a warning label: whether ABE exists in a given preparation is an empirical question you must measure before designing around it.
Third, the cliff reframes substrate quality control. The same hysteresis singularity that gives e-noise its tuning range means a sufficiently noisy substrate stops being excitable at all - not degraded, not noisy-computing, but non-spiking. Organoid variability in channel expression, myelination state, or drug exposure moves cells along exactly this axis. A biocomputing platform that treats noise as a fixed nuisance will silently acquire batches that pass viability stains and fail to compute. The predictability-and-resonance measurements this paper proposes are, read through our lens, incoming substrate QC.
The dual-use and hype angle cuts both ways. The opportunity is real: a two-timescale, two-noise-channel theory is the first clean design space for exploiting biological stochasticity rather than suppressing it. The threat is equally real: stochastic resonance is a narrow-band phenomenon, the separability of the dials is unproven in tissue, and "the brain uses noise" is precisely the kind of slogan that outruns a single-neuron, dimensionless model. Treat this as a design hypothesis with proofs attached, not as a license to stop fighting noise.
The bottom line
Established: in a simplified memristive neuron, with proof under timescale separation, fast excitatory noise modulates excitability and sensitivity - including noise-induced anode break excitability and noise-destroyed excitability past a critical intensity - while slow inhibitory noise drives stochastic spiking gated on ABE, and both contribute to stochastic resonance with the resonant firing rate tracking the signal frequency. Not established: any of this in tissue, in networks, or at biological noise scales; the dials' independence is assumed, not shown. What would confirm: patch-clamp or MEA measurements of ABE and resonance tuning in organoid or cultured neurons, where σ can be manipulated pharmacologically. What would break it: a conductance-based network model showing the ABE criterion fails when slow feedback is dominated by synaptic rather than intrinsic fluctuations.
Frequently asked questions
What is a memristive neuron model?
A reduced conductance-based model in which the excitatory and inhibitory conductances depend on internal state variables with memory (memductances), giving the neuron two feedback pathways: one fast and excitatory, one slow and inhibitory.
What is anode break excitability?
Firing triggered by removing an inhibitory current rather than applying an excitatory one: the slow inhibitory state, dragged below rest, acts as stored deflection that the fast dynamics convert into a spike upon release.
How can noise destroy excitability?
Excitability depends on a folded (reverse-N-shaped) fast current profile. Enough fast-pathway noise smooths the fold away; at a critical intensity the fold vanishes in a hysteresis singularity and the neuron stops being excitable.
What is stochastic resonance here?
A subthreshold periodic input becomes detectable when intrinsic noise is neither too weak nor too strong; in the model, SNR peaks at intermediate noise in both pathways and the resonant firing rate matches the 12.5 Hz signal frequency.
Does this apply to real neurons or organoids?
Not yet demonstrated. The proofs assume a single compartment and timescale separation, the simulations use dimensionless parameters, and the Hodgkin-Huxley check is illustrative. Tissue and network tests are future work.
What is the actionable idea for closed-loop interfaces?
Calibrate stimulation to the substrate's measured noise: exploit resonance by driving below deterministic threshold, and test whether inhibitory-then-release waveforms evoke spikes with less injected charge than excitatory pulses.
References
- Koo, S., Lee, J. G., and Sepulchre, R. Neuromodulation Through Intrinsic Noise. arXiv:2611.13780. 2026. https://arxiv.org/abs/2611.13780. Accessed 2026-11-12.
- Burger, T. S. J., Shahhosseini, A., and Sepulchre, R. A memristive model of spatio-temporal excitability. Proc. IEEE Conf. Decision Control, pp. 13-18. 2025. Citation as given in the reference list of the primary source. Accessed 2026-11-12.