Rhythm by event propagation, not coupled oscillators
Most engineering of rhythmic neural circuits borrows the language of coupled oscillators, and then pays for it in fine-tuning. Robin Wroblowski and Rodolphe Sepulchre of KU Leuven and Cambridge take the opposite bet: build the central pattern generator from excitable nodes that do nothing on their own, let inhibition-triggered rebound propagate events around a ring, and the global rhythm becomes a sum of local delays that can be tuned one node at a time. Their simulations back the bet, an excitable four-node ring keeps its rhythm under parameter heterogeneity that collapses the equivalent oscillatory network.
Source: Neuromorphic design of a rhythmic network, arXiv:2610.39762, October 2026, submitted to an IOP journal. Primary source. Read: the full 23-page PDF including the event describing function and event phase response curve analyses, the heterogeneity benchmark, the gait retuning examples, and Appendix A with the neuron model.
What the work claims
This is a design-and-analysis paper, not an experimental one; every result is a Hodgkin-Huxley-type simulation of small rings, and it should be weighted accordingly. The claim is that the classical oscillator toolkit, describing functions and phase response curves built for autonomous limit cycles, has a legitimate excitable counterpart, and that this counterpart is a better design framework. The two instruments are the event describing function (eDF), which maps an input event period to the steady-state delay a node inserts, and the event phase response curve (ePRC), which maps a perturbing event's phase offset to the timing shift it induces. Because an excitable node returns to a stable resting state between events, both curves can be measured from a single transient response once the network period exceeds the node's reset time, and the global period of a ring reduces to the algebraic condition that the relative nodal delays sum to one.1
The numbers that carry it: for the paper's reference neuron, the eDF has a minimum input period of 14 ms, an asymptotic delay of 11.66 ms, and a reset time of about 26 ms. In the heterogeneity benchmark, 400 trials at each level of Gaussian parameter variation on the sodium and potassium conductances and the synaptic time constant show the oscillatory network losing its sequential rhythm quickly above 10 percent heterogeneity, while the excitable ring keeps a valid sequential pattern across a much wider range. And in an entrainment test on a ring with a 60.9 ms period, an excitatory perturbation train with a period 4 ms shorter than the network's locks the node at a phase offset of 9.13 ms, matching the ePRC prediction.
How it works
The building block is a neuron modelled with two bio-inspired internal currents on top of the classical fast-inward, slow-outward pair: a low-threshold inward current with slow de-inactivation that supplies post-inhibitory rebound, inspired by the T-type calcium current, and an A-type potassium current whose slow inactivation sets the latency of spike generation. In a unidirectional inhibitory ring, each node sits at rest until the previous node's spike hyperpolarises it; release from inhibition triggers a rebound spike, which inhibits the next node. The chain is only as rhythmic as the sum of its nodal delays, the total period is literally the sum of per-node delays, and each delay is set by exactly two knobs. The external knob is the synaptic decay time constant, which tunes the delay roughly linearly; the internal knob is the A-type conductance, which tunes it exponentially over a wider range. Neither knob changes whether the node is excitable, so tuning timing does not destabilise the mechanism that produces spikes.
The eDF is what makes this a design discipline rather than a hope. Swept over input periods, it shows three regimes: below a minimum period the node cannot follow one-to-one and either slips into higher-order locking or stays silenced under sustained inhibition; above the reset time the delay converges to a constant baseline set only by the last isolated event, because the neuron's fading memory erases earlier history. Operating in that fully-reset regime is what decouples local behaviour from global period and makes each node independently characterisable. The ePRC plays the same role for external interaction: for the paper's model, excitatory perturbations always advance the next event and inhibitory perturbations always delay it, with no zero crossings, the signature of Type 1 excitability, and the peak-to-peak amplitude of the ePRC sets the exact window of period mismatch over which an external rhythm can entrain the node, about 5.5 ms for the reference synapse strength.
The payoff demonstrations are gait retunings on a four-limb ring: a walking phase profile of equal quarter phases is reshaped into a galloping profile of 0.1, 0.4, 0.1, 0.4 by moving the A-type conductance of two nodes from 1.2 to 0.0 and two to 1.45, with small synaptic adjustments, either holding the global period fixed or holding two nodes' delays fixed while the period re-solves itself. The same local-isolation property shows up in a single-synapse experiment: lengthening one synaptic time constant from 3.2 to 9.5 ms at a fixed time stretches only that one inter-node delay in the excitable ring, whereas the same manipulation in the oscillatory network, where 3.2 to 4 ms sufficed to break the rhythm, scrambles every effective delay in the network.
Where a skeptic should push
The most load-bearing assumption is sparsity. Every clean result, the single-transient measurement of the eDF, the decoupling of local delays, the algebraic period condition, requires events separated enough that each node fully resets. The paper is explicit that this rests on the inhibitory-rebound architecture and that the design philosophy is to make the nodal delay sensitive to few cellular parameters and insensitive to most. But dense, near-threshold, or bursty operation, which is where interesting neural computation usually lives, breaks the sparse-event premise, and the authors defer bursting networks to future work. A skeptic should also note that the headline robustness comparison pits the minimal excitable ring against what the authors call the most robust oscillatory configuration; the gap, while large, is the best case for the incumbents.
Second, the sign regularity of the ePRC, excitation always advances, inhibition always delays, is stated by the authors to be phenomenological, a property of this neuron and synapse model rather than a theorem, and they flag that specific parameter regimes can invert it. Third, all evidence is simulation of a four-node ring with an idealised two-current neuron; there is no hardware, no noise floor analysis beyond parameter heterogeneity, and no demonstration above ten nodes. The 400-trial robustness curve comes with confidence intervals, which is good practice, but it is a mean over one model family. None of this kills the framework; it bounds where its promises are cashable.
Rhythm design on excitable organoid tissue
For organoid intelligence the paper's premise is the field's premise. The neurons in a cortical organoid are, individually, excitable elements with no intrinsic rhythm; the oscillations and bursts that MEA recordings show are network products. What the field has lacked is a language for that situation that is not borrowed from coupled-oscillator theory, which quietly assumes every node is a little clock and then struggles to explain why real tissue tolerates enormous cell-to-cell variability. Wroblowski and Sepulchre supply that language: treat the tissue as a network of event-propagating nodes, measure an eDF and an ePRC per node or per local subnetwork, and the rhythm becomes an algebraically tractable object. The practical research programme writes itself: map an organoid's dominant rhythmic motifs as event chains, perturb single units with optogenetic or electrical pulses at swept phases to measure ePRCs, and use those curves to predict, rather than search for, the stimulation parameters that entrain the network to a desired period.
The non-obvious implication is the two-knob result. Synaptic time constants and A-type potassium conductances are chemically and genetically addressable in real tissue; drugs that block A-type currents exist, and neuromodulators shift synaptic decay times. The paper says these knobs have different gain profiles, linear versus exponential, and that covarying them adjusts a node's sensitivity to incoming events while leaving its nominal timing untouched. For a wetware engineer that is a control channel: you could in principle make an organoid subcircuit more or less responsive to a given input pattern without changing the pattern it is currently producing, which is close to an attractor-selection mechanism for living computers. The entrainment arithmetic sharpens the same point: an external periodic drive can only capture a node if the period mismatch fits inside the ePRC amplitude, so the ePRC is literally a certificate of controllability for closed-loop stimulation, something the organoid field currently tunes by hand.
The threat is that the framework's virtues are exactly what tissue does not grant. Independent per-node tuning assumes you can address one node's A-type conductance without touching its neighbours, which no current modality does; sparse 1:1 event propagation assumes a duty cycle far below what dense recurrent tissue exhibits; and the robustness claim assumes heterogeneity is static, while organoids drift developmentally on a timescale of days. There is also a subtler warning. If organoid rhythms are event-propagating chains rather than coupled oscillations, then decoding and modelling work that fits oscillator phase models to organoid recordings is fitting the wrong object, and its phase-locking predictions will fail precisely under the heterogeneity that tissue guarantees. The opportunity is a design science for organoid rhythms; the threat is discovering that the design variables are not independently accessible and that the beautiful local analysis remains a silicon-adjacent idealisation.
The bottom line
Established, in simulation: excitable rings generate robust sequential rhythms under heterogeneity that breaks oscillatory rings; local synaptic and cellular parameters tune global period and phase profile with predictable, isolated effects; and measured ePRCs predict entrainment windows quantitatively, 9.13 ms against a predicted zero crossing, with a capture window set by the ePRC amplitude. Not established: anything on hardware, in bursting regimes, at tissue scale, or under realistic synaptic noise. The confirming result for organoid intelligence would be an experiment that measures eDF and ePRC curves for organoid subnetworks and uses them to entrain a tissue rhythm on the first predicted parameter set rather than by scan. The breaking result would be evidence that organoid rhythmicity is dominated by dense bursting that never enters the sparse 1:1 regime, in which case this framework becomes a boundary marker: the regime where rhythm is designable, and where it is not.
Frequently asked questions
What is a central pattern generator?
A neural circuit that produces rhythmic motor commands, such as those driving locomotion or respiration, without needing a rhythmic input. Classically they are modelled as networks of coupled oscillators; this paper models them as rings of excitable nodes that pass events along.
What is the event describing function?
A curve that records the steady-state delay a node inserts between an incoming periodic event train and its outgoing spikes, as a function of the input period. It plays the role that a describing function plays for oscillators, but for excitable nodes that return to rest between events.
What is post-inhibitory rebound?
The property that a neuron fires not when it is inhibited but when the inhibition is released. In the paper's rings, inhibitory synapses are therefore not suppressive but propulsive: each node's spike is triggered by the end of the previous node's inhibition of it.
Why are excitable nodes more robust to heterogeneity than oscillators?
Oscillatory nodes bring intrinsic frequencies that must be reconciled globally, so parameter variation breaks phase-locking; in the paper's benchmark the oscillatory network degrades quickly beyond 10 percent heterogeneity. Excitable nodes have no intrinsic frequency, so variation in one node shifts only its own local delay, and the rhythm survives far larger mismatch.
What does the event phase response curve predict?
The steady timing shift caused by an additional periodic input at a given phase offset. Its peak amplitude bounds the range of external periods that can entrain the node; beyond that window no stable locking exists, which turns entrainment design into curve reading.
Could this framework control an organoid's rhythm?
Possibly in the sparse-event regime. The required steps, measuring per-subnetwork response curves and choosing stimulation phases inside the entrainment window, are feasible with current optogenetic and multielectrode tools. What is missing is independent addressability of single nodes and evidence that organoid rhythms operate in the 1:1 sparse regime the analysis assumes.
References
- R Wroblowski, R Sepulchre. Neuromorphic design of a rhythmic network. arXiv:2610.39762. 2026. https://arxiv.org/abs/2610.39762. Accessed 2026-11-21.