When a spiking network writes a border into its own synapses

What the work claims

Know the genre before weighting the result. This is a computational modelling paper by a single independent researcher, posted to arXiv and not peer reviewed.¹ It reports no biological data. It never mentions organoids, cell culture or microelectrode arrays. Every claim below about living tissue is my extrapolation, not the author's, and I will mark where that extrapolation is load bearing.

The setup is deliberately austere. A grid of 200 by 100 excitatory and inhibitory neuron pairs, 40,000 neurons and roughly 7.9 million synapses, runs in discrete synchronous time steps with binary spikes and a one step refractory period. Membrane potential resets to zero every step, so individual neurons have no temporal integration at all: every trace of the past lives in synaptic weights and firing thresholds. Excitatory cells reach a radius of 9 grid points, inhibitory connections only 5. That radius asymmetry alone produces a functional Mexican hat profile, short range excitation surrounded by relatively longer reach inhibition, without any distance dependent weighting of synaptic strengths. All weights start spatially uniform.

Drive a small disk of excitatory cells at a fixed period and the network organises itself into annular wave packets, a bright excitatory ring trailed by inhibition, moving outward at constant speed. Stop the stimulation and the inner rings vanish immediately while the outermost packet keeps going on its own. The plasticity rule, a multiplicative weight dependent variant of spike timing dependent plasticity that the author calls WSTDP, strengthens synapses pointing along the direction of travel and weakens the reverse, so the wave lays down its own track.

Then the result that earns the paper its keep. Two sources, one at x equals 50 and one at x equals 150, driven at the same rate and phase, produce waves that meet and annihilate at the midline. Silence the left source, and the right wave propagates freely across empty territory and still halts abruptly at x equals 100, with nothing there to stop it. Offset the two sources by three time steps and the boundary moves off the midline, and the surviving wave stops at the shifted position instead. The collision line is not an event. It is a structure, stored in the weights, that outlives the thing that made it.

How it works

Two ingredients do the work, and the paper credits the wrong one more loudly than it credits the right one.

The advertised ingredient is WSTDP. Spike timing dependent plasticity is a Hebbian rule in which a synapse strengthens when the presynaptic cell fires shortly before the postsynaptic cell and weakens in the opposite order. Here the update is multiplicative, meaning the change depends on the current weight, so weights settle at a fixed point set by the ratio of potentiation to depression events rather than slamming into a hard ceiling.

The unadvertised ingredient is the input normalisation, and it deserves top billing. Each neuron divides its incoming drive by the sum of its own excitatory weights. Consider what that does. Under the multiplicative plasticity rule alone, every synapse would relax to its own fixed point independently; nothing makes synapses compete for anything. Dividing by the total excitatory weight converts synaptic strength into a zero sum resource within each neuron, and that competition is the actual engine behind both headline phenomena. The author states as much for directional locking: once a direction is learned, the reverse direction fails because normalisation dilutes the few surviving backward synapses below threshold. The persistent boundary is the same story arranged in space. The honest one line summary is that these results are consequences of synaptic competition imposed by normalisation, with the plasticity rule supplying only the sign.

The territorial dynamics follow naturally. When the two sources run at different rates, the faster one delivers more potentiation events and its territory expands. At periods of 7 against 6 the boundary drifts slowly and has not settled by the end of the run. At 7 against 5 the faster source takes over decisively, leaving only a small wedge where waves from both sources happen to travel the same way and reinforce rather than cancel. The boundary is therefore semi-persistent: robust against weak perturbation, rewritable under sustained drive, which the author reasonably likens to metaplasticity.

Where a skeptic should push

The single most load bearing assumption is that the normalisation is exact, instantaneous, and global across a neuron's excitatory inputs. Every biologically available implementation of divisive normalisation is approximate, delayed, noisy and at best confined to a dendritic compartment. The author proposes a dendritic scheme in which each branch junction attenuates its subtrees by their relative weight sums using only local information, and candidly concedes it conflicts with real neurons, where potential summates toward the soma rather than staying bounded. I would not press that concession hard, because divisive normalisation is a well supported cortical computation and the obvious reply is to implement it with shunting inhibition instead. The sharper question is one the paper never asks and that an afternoon of simulation would answer: do directional locking and boundary persistence survive a normalisation signal that is low pass filtered, noisy, or compartment local? If the boundary dissolves under a realistic time constant, the result is an artefact of an idealised global operation.

Separate the demonstrated from the asserted. Demonstrated, as single realisation existence proofs inside the model: wave formation from uniform weights, autonomous propagation after stimulation ends, directional locking, and above all the silencing control, which is properly conceived and is what distinguishes a stored structure from a transient collision. Asserted: that boundary position encodes relative phase. One non-zero offset does not establish a code. Monotonicity across the full range of offsets would. Likewise the claim that boundary position is proportional to relative potentiation flux rests on two conditions with no fit and no regression, and in one of those two the boundary never stabilised.

The documentation has real defects. The threshold parameters differ between the methods section and the results subsection that studies thresholds, which is awkward precisely because that subsection's finding is a claim about the maximum threshold. The radius sweep timings are described by the author himself as visual approximations, so the specific step counts should not be quoted as measurements by anyone, including me. What survives that imprecision is the qualitative asymmetry, and it is mechanistically informative: setting the inhibitory reach onto excitatory cells equal to the excitatory reach abolishes waves entirely, whereas the reciprocal manipulation merely slows them. I found no statement of random seeds, repeat runs, or code and data availability anywhere in the paper, which for work whose entire evidence base is simulated trajectories is close to disqualifying for a strong claim. The paper does carry a conflict of interest statement declaring none.

One more caution about interpretation. With synchronous updates, binary spikes and no membrane integration, the substrate is closer to a probabilistic cellular automaton than to a spiking neural network. That does not invalidate the phenomenon, but it means nothing quantitative about speeds, timescales or emergence times should be carried over to tissue. The global synchronous oscillation seen when inhibition is weakened, which the author calls analogous to epileptic discharge, is also the textbook signature of a synchronous update rule, so treat that particular analogy as rhetoric rather than finding. Here, though, the author deserves credit rather than criticism: the paper carries an explicit limitations subsection naming precisely these issues, including discrete synchronous time, one spike per step with no bursting, the membrane reset that eliminates leaky integration, and stimulation applied directly rather than synaptically. An author who lists the objections to his own model has earned the right to have them read as declared scope rather than as concealment.

I will note one disagreement within my own review process. An independent reviewer of this analysis argued the paper misuses the word soliton, since true solitons pass through one another intact while these annihilate. Checking the source, that criticism does not land: the author explicitly identifies them as dissipative solitons and explicitly notes that annihilation on collision is characteristic of that class in non-equilibrium systems such as reaction diffusion media and optical cavities. The terminology is careful. I record the exchange because it is exactly the kind of plausible objection that survives if nobody opens the paper.

What this means for organoid intelligence

The obvious reading is that this hands you a way to store information in tissue you cannot address at single cell resolution. That reading is roughly right, and the reason it is right is not the one most people would give.

The licensing fact is that the network begins with spatially uniform weights. All of the structure, the wave, its direction, the boundary, is imported from the statistics of the input rather than from the initial wiring. The stored variable is therefore a collective coordinate: a location determined by drive statistics in which no individual synapse is load bearing and no individual synapse needs to be found, read, or even to survive. That is precisely the property a developmentally random, heterogeneous, partly dying substrate requires. Coarse electrodes suffice not because the effect is large but because the encoding is collective, and collective encodings tolerate exactly the substrate variability that currently makes organoids so hard to use. Concretely: two stimulation sites on a microelectrode array, driven at fixed rates, would place a boundary whose position is written by their phase offset and rewritten by their frequency ratio.

Two corrections to the enthusiasm, both grounded in the paper's own mechanisms. First, this is not addressable memory. It is one scalar per source pair, with precision bounded by wavelength and electrode pitch. Call it an analog register. Second, and more consequentially, the paper contains no readout. The boundary is visible to someone inspecting the entire weight field, which no experimenter can do in tissue. Reading it electrically means driving one source and detecting where propagation stops, and the paper's own threshold result says that sustained drive erodes the boundary. Reading is therefore destructive. What this actually specifies is a read-disturb-limited register requiring refresh, architecturally closer to DRAM than to flash, with a read budget and a measurable parameter: boundary drift per probe. That is a considerably more useful engineering specification than a vague claim about mesoscale memory, and it falls straight out of the source.

Now the threat, which I think is the most valuable thing in this paper and is buried in its control condition. Weaken inhibition and the network stops computing and starts oscillating globally, concentric rings filling the entire surface, no spatial confinement, no waves. Get the inhibitory radius wrong in the other direction and you get fibrillation, or nothing at all. So the window of excitatory and inhibitory balance that permits structured computation is narrow, and it is set by connection geometry. Now recall what neural organoids on arrays characteristically do: they burst, synchronously, across the whole culture. The paper's pathological regime is the default observed behaviour of the substrate the field wants to compute with. That inverts the paper's relevance. Its most defensible contribution to organoid intelligence is not a programming manual but a diagnostic criterion: a mechanistic account of why cultures burst instead of computing, and therefore a test of whether a given preparation sits in the computational regime at all. Every proposal above is conditional on passing that test, and geometry is the one parameter an organoid experimenter cannot specify, because it emerges from self organised development. That is an argument for closed loop homeostatic control of excitability as a prerequisite for training, not an argument that training is hopeless.

There is a quieter dual-use style hazard too. The frequency dominance result is an unintended blueprint for stimulation protocols that silently overwrite previously learned state. Any long running organoid experiment that changes its stimulation rate partway through may be erasing what it thinks it is measuring, and would have no way of knowing.

Finally, the substrate obsolescence question, asked honestly. Forty thousand neurons of this behaviour simulate at roughly eleven seconds per thousand steps on one laptop processor. If this class of dynamics is what wet tissue offers, silicon already does it faster, more reproducibly and without an incubator. The case for the biological substrate cannot rest on phenomena a laptop reproduces. It has to rest on something the simulation is not capturing, which is precisely why the honest framing of this paper is diagnostic rather than aspirational.

The bottom line

Established, within the model: a recurrent excitatory and inhibitory network with purely local plasticity and uniform initial weights can generate self sustaining travelling waves and can freeze the locus where two wave territories meet into its synaptic weights, such that the boundary survives removal of one source. The silencing control is real evidence and the paper deserves credit for running it.

Hypothesis, not result: that boundary position constitutes a code for relative phase or rate, that the effect survives biologically realistic normalisation, and that any of it transfers to living tissue. The paper claims none of that last part, and I want to be clear that the transposition is mine.

What would confirm it: a systematic sweep of phase offsets showing monotonic boundary displacement, with repeats and reported seeds; an ablation replacing exact global normalisation with a lagged, noisy, compartment local approximation and showing the boundary persists; and released code. What would break it: boundary persistence disappearing under realistic normalisation, or under asynchronous updating with heterogeneous initial weights, either of which would identify the effect as an artefact of the idealisation rather than a property of recurrent plastic networks. In tissue, the decisive experiment is simpler than it sounds. Drive two well separated electrodes on a mature culture at different fixed rates, stop one, and test whether evoked propagation from the survivor halts at a reproducible location that shifts with the phase offset you imposed. A negative result there would be as publishable as a positive one, and considerably more likely.