Research analysis · Computational neuroscience

The firing-rate contract: what one number in an organoid paper actually promises

Every organoid electrophysiology paper reports a firing rate, usually as one smoothed trace. A new 39-page review from the Federal University of Santa Catarina lays out something the field glosses: firing rate is three different quantities that agree only under assumptions, and the choice among them is a modeling decision that changes the science. The same review's worked examples, recurrent bistability and synaptic-depression rhythms, map directly onto the activity states organoid researchers see and argue about.

Source: Mean-field theory of neural networks: from spikes to firing rates, Carvalho, Haas, Caprioti, Delgado, Kuznetsov, Schweitzer, Vasconcelos, Rocha and Girardi-Schappo, arXiv:2611.17512, 2026. Primary source. Read: full PDF, verified inline.

What the work claims

This is a review and teaching monograph, not a new primary result, and it is explicit about its genre: it formalizes definitions, works through canonical derivations, and surveys applications. Its organizing claim is that firing-rate models are the legitimate intermediate description between single spikes and whole-brain dynamics, justified by a thermodynamic analogy: just as temperature is a meaningful macroscopic variable that no single molecule possesses, a population firing rate is a macroscopic variable that emerges from, and is not reducible to, any individual spike. The review's real contribution for practitioners is taxonomic. It separates three distinct constructions that all travel under the name firing rate: a temporal average over one neuron's spike train, a population average across neurons at an instant, and an average across repeated trials of the same stimulus. They coincide only when activity is stationary, homogeneous, and ergodic, conditions that most real preparations, and especially small in vitro networks, do not meet.1

How it works

On the temporal side, the review builds from the stationary rate, spike count divided by observation time, through binned counts, to a general definition: convolve the spike train with a normalized kernel. Rectangular, exponential, alpha, and Gaussian kernels give different low-pass filters, and the choice of width is a trade between temporal resolution and smoothing. Two points the authors insist on carry practical weight. First, the convolution must be applied to the spike train, not to raw membrane voltage. Second, the operation is irreversible: spike times cannot be recovered from a rate, so a reported rate is already a theory-laden reduction. A boxed argument gives the operation a biological interpreter: synaptic integration of incoming spikes is itself approximately a convolution with the alpha-shaped postsynaptic current, so postsynaptic neurons naturally read a temporally averaged rate when their inputs are dense.1

On the population side, the review derives mean-field descriptions of recurrent excitatory-inhibitory networks. A worked example with fixed synaptic resources shows recurrent self-excitation creating three stationary states: a low state near 0.1 spikes per second, an unstable threshold near 12.3 spikes per second, and a high state near 23.1 spikes per second, for the paper's stated parameters. A transient input that pushes the population past the unstable middle state flips the network into the high state, where it stays after the input ends: a one-bit memory implemented purely by recurrent coupling. Adding activity- dependent synaptic depression, with a resource recovery time constant of 250 ms, destroys the high states; under constant stimulation the population instead settles into an oscillation with a period of roughly 700 ms that exists only while the stimulus lasts. Activity depletes synaptic resources, depletion weakens recurrent excitation, firing falls, resources recover, and the cycle repeats.1

The applications chapter reads, almost point for point, like a list of organoid phenomena. Wilson-Cowan-type rate models reproduce synchronous oscillations and multistability that serve as model seizure states, and cited work shows that weak high-variance or high-frequency stimulation, periodic or stochastic, disrupts the multistability and prevents transitions into seizure states, an effect attributed to selective recruitment of inhibitory neurons. Neural-mass models such as Jansen-Rit generate epileptiform activity and EEG/MEG signatures. Extensions incorporating refractory-period memory effects reproduce avalanche statistics and change the nature of transitions to sustained activity, with scaling characteristic of directed percolation.1

Where a skeptic should push

The most load-bearing assumption is that mean fields exist: that averaging over neurons, time, or trials yields a closed, low-dimensional dynamics. This fails exactly where organoid work lives, in small, heterogeneous, non-stationary networks where single-cell and finite- size effects dominate, and where trial averaging is impossible because no two minutes of recording are the same stimulus repeated. The review is honest that its worked examples use hand-set parameters chosen to make the dynamics visible, and that the place-field code it uses as motivation may be a position-correlated artifact rather than a genuine rate code; the same caution applies to any rate-based interpretation of an organoid's burst statistics.1

A subtler trap is the unit itself. The authors note, correctly, that calling a spike count per second a frequency in Hertz is an abuse: Hertz belongs to periodic waves, and a spike rate is an event density. That pedantry matters operationally. A 10 Hz rhythmic oscillation and 10 spikes per second of Poisson bursting are different physical objects, and analyses that smooth both into one rate trace have already decided they are the same. The review's own seizure example shows what is at stake: the therapeutic variable is variance and spectral content of the stimulation, information that exists only above the rate description.

What the rate contract means for organoid benchmarks

The non-obvious implication is that organoid benchmarking currently runs on an unsigned contract. When a paper reports that an organoid's firing rate rose under stimulation, a skeptical reader cannot tell which of three quantities was measured: a temporal average of one unit, a population average at an instant, or a trial average across repeats, nor which kernel or bin width was applied. The review shows these are not pedantic variants. A population average can look silent while individual rates carry a clear pattern, and a Gaussian kernel needs spikes from the future, which makes a real-time decoder's rate a different object from an offline plot's rate. An organoid-computing field that wants reproducible benchmarks needs to state its averaging sense, kernel, and bin width the way a silicon paper states its precision and clock speed.1

The opportunity is theoretical leverage. Organoids routinely exhibit discrete activity regimes, quiet, bursting, and seizure-like, and the review's recurrent-bistability construction gives the simplest mechanistic account: two stable fixed points separated by an unstable threshold, with inputs switching between them. That reframes a working memory claim for organoids into a testable question: does the preparation hold a high-activity state after the stimulus ends, and is the hold-state destroyed by synaptic depression as the depression model predicts? The same machinery hands closed-loop organoid control a ratified blueprint. The cited seizure results say weak, high-variance or high-frequency stimulation prevents entry into hyperactive states by recruiting inhibition, which is a stimulation protocol any electrode-equipped organoid system can try tomorrow, with the rate-model prediction as a preregistered hypothesis rather than a fishing expedition.

The threat is double-edged. On one side, rate-based reporting can launder artifacts: smoothing choices alone can manufacture apparent task-correlated modulation from drifting burst dynamics, and the irreversibility of the rate operation means the raw evidence is gone once the trace is plotted. On the other side, oscillation is not maturity: the review's synaptic-depression example generates a clean 700 ms rhythm from a depressing recurrent loop with constant input, so an organoid that oscillates has demonstrated a depressing excitatory loop, not necessarily organized circuitry. Treating oscillation as a maturity biomarker, common in the field, is exactly the over-reading this formalism warns against.

The bottom line

Established: the three averaging constructions are mathematically distinct and coincide only under stationarity, homogeneity, and ergodicity; kernel-based temporal averaging is irreversible; recurrent coupling supports bistable persistent states, and synaptic depression converts them into stimulus-bound oscillations. Established as modeling results, with hand-set parameters: the specific state values and periods. Open: how far these mean-field reductions survive in finite, heterogeneous organoid networks. What would confirm the transfer: organoid experiments that preregister the predicted bistable switching and depression-mediated oscillation signatures, with the averaging sense declared. What would break it: systematic failures of rate reductions to predict spike-resolved organoid dynamics despite adequate recording quality, which would push the field back toward spike-level descriptions for anything claiming computation.

Frequently asked questions

What are the three meanings of firing rate?

A temporal average of one neuron's spikes over a window or via kernel convolution; a population average of many neurons' activity at an instant; and a trial average across repeated presentations of the same stimulus. They answer different scientific questions and coincide only under stationarity, homogeneity, and ergodicity.

Why is a Gaussian kernel problematic for real-time decoding?

It assigns weight to spikes before and after the evaluation time, so the offline smoothed rate at time t uses future spikes. A causal decoder must use one-sided kernels such as the exponential, which is a different estimator with different properties.

What does recurrent bistability give you for free?

With fixed synaptic resources, the worked example sustains either a low or a high population state after a transient input crosses the unstable middle state, implementing a one-bit memory from recurrence alone, with no synaptic plasticity required.

How does synaptic depression create oscillations?

Firing consumes synaptic resources, weakened resources reduce recurrent excitation, firing falls, and resources recover on a 250 ms time constant. The feedback loop oscillates at roughly 700 ms period, but only while constant stimulation persists; the rhythm stops when the input ends.

What stimulation prevents seizure-like states in these models?

Cited Wilson-Cowan-based work shows weak, high-variance or high-frequency stimulation, periodic or stochastic, disrupts multistability and prevents transitions into hyperactive states, via selective recruitment of inhibitory neurons rather than excitation.

Is a spike rate in Hertz the same as a brain rhythm in Hertz?

No. The review insists on the distinction: Hertz describes periodic waves, while a spike rate is an event count per second. Ten spikes per second of irregular bursting and a 10 Hz oscillation are different physical objects that a smoothed rate trace can accidentally equate.

References

  1. T. T. A. Carvalho, M. Haas, D. D. Caprioti, L. A. R. Delgado, A. Kuznetsov, H. L. Schweitzer, C. F. Vasconcelos, R. P. Rocha and M. Girardi-Schappo. Mean-field theory of neural networks: from spikes to firing rates. arXiv:2611.17512. 2026. https://arxiv.org/abs/2611.17512. Accessed 2026-11-16.