Neurons in visual cortex receive fairly steady inputs, yet their spike counts vary enormously from trial to trial. The standard explanation is the spike threshold. Small, constant fluctuations in membrane voltage produce no spikes when the cell sits far below threshold, produce wildly variable spiking near threshold, and so turn constant "additive" input noise into output variance that grows with the firing rate ("multiplicative" noise). That account treats the threshold as a fixed line. Real cortical thresholds are not fixed: in pyramidal neurons the threshold rises to follow slow depolarizations, because sodium channels inactivate, and so it acts as a high-pass filter whose integration window shrinks as the cell is driven harder.
We propose that this changes the story in a specific, testable way. A voltage-tracking threshold should convert fast input noise into multiplicative output variance much as a fixed threshold does. It should largely cancel slow input noise, and the cancellation should get stronger as mean drive rises. If so, whether a neuron's variability looks additive or multiplicative is not a fixed property of rectification. It depends jointly on the timescale of the noise and the cell's operating point. Part of the shift from multiplicative to additive shared variability seen at high contrast, which is currently attributed to recurrent circuit saturation, could then arise inside single neurons.
Core Claim
Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex motivates the question and supplies the reference model. It showed in cat V1 that generator-potential variance grows by less than a factor of 4 across stimuli while firing-rate variance grows by nearly 100. A static power-law rectification, $R = k[V - V_{\text{thresh}}]^n$, applied to Gaussian voltage fluctuations accounts for this. Here $R$ is the firing rate, $V$ the membrane potential, $k$ a gain, and $n$ an exponent. The paper explicitly leaves the threshold static and leaves the function of the amplification open.
Enhanced Sensitivity to Rapid Input Fluctuations by Nonlinear Threshold Dynamics in Neocortical Pyramidal Neurons supplies the method. Its inactivating generalized integrate-and-fire model (iGIF, a simple spiking model whose threshold is coupled nonlinearly to voltage) predicts spike times of mouse L5 pyramidal cells with 75–83% accuracy at ±4 ms. It also shows the effective integration timescale shrinking from about 27 ms to 5 ms with depolarization. That paper never repeats trials to measure how output variance scales with the mean.
Our claim is that, for a neuron with fluctuating additive input of correlation time $\tau_\eta$, trial-to-trial count variance decomposes as
$$\mathrm{Var}[r] = \sigma_a^2(\tau_\eta,\mu) + \sigma_g^2(\tau_\eta,\mu)\,\bar r(\mu)^2,$$
where $\bar r(\mu)$ is the mean rate at mean drive $\mu$, $\sigma_a^2$ is an additive term, and $\sigma_g^2$ is a multiplicative gain term. This is the affine form used by Circuit-motivated generalized affine models characterize stimulus-dependent visual cortical shared variability. Under a static threshold, $\sigma_g^2$ depends only weakly on $\tau_\eta$. Under an iGIF threshold, we predict that $\sigma_g^2$ falls steeply once $\tau_\eta$ exceeds the threshold-tracking timescale, and falls further as $\mu$ increases. The new element is the prediction that the multiplicative fraction is set by noise timescale and threshold kinetics, together with the experiments that can falsify it.
Research Program
The first stage is analysis and simulation of single neurons. We take iGIF parameters in the range fitted by Mensi et al. and drive the model with a frozen, repeated stimulus-like input of varying mean. We add independent Ornstein–Uhlenbeck noise, a standard random process with a tunable correlation time $\tau_\eta$, sweeping it from 1 to 500 ms at matched voltage variance. Counts are measured in 5–20 ms windows to match Carandini's analysis. For each condition we fit the affine decomposition and record the multiplicative fraction $\sigma_g^2\bar r^2/\mathrm{Var}[r]$. We also derive an approximate prediction by treating the threshold as a low-pass copy of $V$ and passing the residual through Gaussian rectification. That approximation predicts a crossover near the threshold's tracking timescale, and the simulations check where it breaks.
The second stage asks whether the single-cell effect reproduces the high-contrast regime shift. Xia et al. obtain multiplicative variability at low rates and additive variability at high contrast from recurrent saturation in a static power-law rate network. We replace the static nonlinearity with iGIF units, or with an iGIF-derived, state-dependent transfer function. We then ask how much of the shift survives once recurrent suppression is removed by setting recurrent weights to zero. This makes Xia et al. a constraint rather than a foil: if their circuit account already captures everything, our single-cell contribution should be negligible.
The third stage is an in vitro test. Using the somatic current-injection protocol of Mensi et al. on L5 pyramidal cells, we inject repeated frozen signals plus fresh noise whose spectrum is either slow or fast at equal variance, across several mean levels. We measure the same affine decomposition. The protocol adds trial repetition and controlled noise spectra to an established preparation.
Baselines
The naive reference is the static Gaussian-rectification model of Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex, fitted to the same subthreshold statistics. Its only reported results are the variance-growth factors (less than 4 for voltage, nearly 100 for rate); no reported result found on Lacuna for its prediction of variance–mean scaling under spectrally controlled noise. Beating it means predicting the held-out variance–mean curve better at long $\tau_\eta$ while matching it at short $\tau_\eta$, where both models should agree.
The strongest existing single-cell model is iGIF itself, as reported in Enhanced Sensitivity to Rapid Input Fluctuations by Nonlinear Threshold Dynamics in Neocortical Pyramidal Neurons, with 75–83% spike-timing accuracy at ±4 ms. Because our hypothesis uses iGIF, "beating" here means something specific. iGIF fitted only to single-trial spike timing should predict the trial-to-trial variance decomposition without being refit to it. A generalized integrate-and-fire model with spike-triggered adaptation but no voltage-coupled threshold should not.
The key ablation removes the voltage coupling while keeping everything else, including the fitted adaptation currents, the subthreshold dynamics, and the parameter-fitting budget. All models are fit by the same maximum-likelihood procedure on the same training traces and evaluated on held-out repeated trials.
The circuit baseline is the recurrent power-law network of Circuit-motivated generalized affine models characterize stimulus-dependent visual cortical shared variability. It is matched in connectivity and in input variance, and differs only in the unit nonlinearity. Specific numerical variance results for that model are not reported in the file.
Why not just…?
Why not just put a linear high-pass filter in front of a static threshold? This is the most serious threat. A fixed filter would also cancel slow noise. The difference we predict is state dependence. Mensi et al. show the effective timescale moving from about 27 ms to 5 ms with mean drive, and a fixed filter cannot make $\sigma_g^2$ fall with $\mu$ at fixed $\tau_\eta$. We will fit the best fixed filter at the middle operating point and test it at low and high means. If it matches the dynamic threshold across the range, the claim reduces to "filtering matters," and we will report it that way.
Why not just attribute the additive shift to recurrent saturation, as Xia et al. do? That account may be complete. The zero-recurrence condition in stage two settles it. If iGIF units without recurrence show no shift toward additive variability at high drive, the single-cell contribution is negligible, and the proposal's circuit claim fails while the single-cell claim may still stand.
Why not just refit the power-law exponent $n$ at each operating point? A drive-dependent exponent can mimic many variance–mean curves, but it cannot depend on noise spectrum at fixed variance. The slow-versus-fast injection contrast in stage three is exactly the condition under which a refit-exponent model must fail if our hypothesis is right.
Decisive Checks
Three results would support the claim. First, in simulation, the multiplicative fraction under iGIF should drop by a clear margin between fast and slow noise at equal voltage variance, while the static and adaptation-only ablations show little dependence on $\tau_\eta$. Second, the multiplicative fraction should decrease with mean drive under slow noise, in a way the matched fixed-filter model cannot reproduce. Third, in recorded L5 cells, the slow-versus-fast spectral contrast should change the affine decomposition in the direction iGIF predicts from timing-only fits.
Either of two results would falsify it. One is spectral independence of the variance decomposition in recorded cells. The other is a matched linear filter performing as well as the dynamic threshold across operating points.
A positive outcome would also give a candidate answer to Carandini's open question about the function of amplification. Cortical neurons would let fast, potentially informative fluctuations become rate-scaled variability while removing slow, shared drifts. The same treatment could later be applied to spiking-network thresholds such as Threshold Modulation for Online Test-Time Adaptation of Spiking Neural Networks, which adjusts thresholds only from batch statistics. We leave that extension outside the present scope.
References
- Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex
- Enhanced Sensitivity to Rapid Input Fluctuations by Nonlinear Threshold Dynamics in Neocortical Pyramidal Neurons
- Circuit-motivated generalized affine models characterize stimulus-dependent visual cortical shared variability
- Threshold Modulation for Online Test-Time Adaptation of Spiking Neural Networks