Bump wandering gives slow MEA drift a testable signature

What the work claims

This is a theory paper: a stochastic neural field model, with the mathematics done in the high-gain limit where firing rates become step functions, and every result verified by numerical simulation of the model itself. The claim has three layers¹. First, splitting inhibition into two populations, fast perisomatic parvalbumin (PV) cells and broader-footprint somatostatin (SST) cells, changes both whether persistent activity bumps exist and how they lose stability. Second, stability decomposes into two clean modes: a shifting mode, in which the bump translates and, past the boundary, travels; and a scaling mode, in which the bump's width oscillates and can annihilate entirely. Which mode kills the memory is determined by inhibitory timescales. Third, for a stable bump, noise drives the bump position as unbiased diffusion: position variance grows linearly in time at a rate the authors derive analytically and show is slowed by broadening SST connectivity.

The bold move is replacing the single homogeneous inhibitory population used in forty years of attractor models with two populations wired as cortical measurements suggest: PV neurons inhibit each other and pyramidal cells locally; SST neurons inhibit PV cells and spread their influence over a wider spatial footprint; SST self-inhibition and PV-to-SST coupling are weak or absent. The dynamical consequences, derived and then simulated, are that SST recruitment thresholds constrain bump existence more strongly than PV thresholds, that weak inhibitory-to-inhibitory connections reshape stability nonmonotonically, and that a wider SST footprint both expands the stable region and reduces the diffusion rate of the memory representation.

How it works

The model lives on a ring representing a continuous feature, such as head direction or remembered location, with excitatory, PV, and SST synaptic input fields coupled through spatially decaying kernels. In the Heaviside firing-rate limit a bump is characterised entirely by its threshold-crossing interfaces, which reduces existence and stability to algebra. Linearising about a stationary bump separates perturbations into shifting and scaling symmetry classes, so the largest eigenvalue can be read separately in each subspace: a positive shifting eigenvalue means drift toward a traveling wave, a positive scaling eigenvalue means width oscillation and eventual collapse. Sweeping the inhibitory timescales shows the organising fact: inhibition faster than excitation stabilises, and as either inhibitory timescale lengthens, the bump crosses one of the two boundaries, with slow SST drive favouring drift instability and matched slow timescales favouring scaling annihilation.

For the wandering, the authors project weak additive noise onto the bump's translational zero mode using a solvability condition, which yields an effective diffusion equation with variance equal to a derived constant times time. The constant depends on interface slopes, population timescales, and the adjoint null vector localised at the six threshold interfaces, and the noise contributions of the three populations add as independent terms. Numerically, across 10,000 simulations per parameter set on a 16,384-point ring with 0.25 ms steps, position variance grows linearly with slopes matching the prediction. Increasing the SST spatial footprint from the baseline toward ten times broader lowers the predicted and simulated diffusion rate, so wider long-range inhibition means a more precisely held memory.

Where a skeptic should push

The most load-bearing assumption is that real cortical tissue sits in the parameter regime where this model's predictions apply at all. Every number here is internal to the model: the firing rates are step functions, synapses are static, the bump is a single contiguous region, the noise is weak, white, and independent across populations, and the PV/SST architecture is imposed rather than emergent. VIP interneurons, the third major subtype and a prominent disinhibitory pathway, are absent; short-term synaptic plasticity, which could move the stability boundaries as a function of network state, is absent. The model was not fit to any recording. That does not make the results empty, because the mechanism-level claims, such as the existence of separate shifting and scaling instability routes, are qualitative enough to survive detail, but every quantitative rate should be treated as illustrative until someone measures one.

Two softer caveats matter for the use I am about to make of it. The diffusion law is derived in the small-noise, near-stationary limit; the paper itself notes that correlated noise across populations adds cross-terms it does not include, and strong perturbations can knock the system out of the linear regime entirely. And the whole framework presumes the tissue actually maintains bump-like persistent states; whether organoid or slice preparations on an array hold anything describable as a continuous-attractor memory is an open empirical question, not something this paper touches.

Interneuron structure and the wandering-rate readout

The non-obvious implication for array work concerns attribution of slow non-stationarity. Every chronic MEA recording shows it: the centroid of sustained activity creeps over hours, burst patterns wander, baselines drift. Operators assume instrument drift and re-reference; biologists assume the tissue changed and re-baseline. The paper says these are statistically different processes. Instrument drift, being deterministic or driven by slow environmental fluctuations, does not in general produce position variance that grows linearly and without bias; tissue state wandering under this theory does, at a rate set by circuit parameters. So the variance-versus-time curve of a tracked activity centroid is a diagnostic: linear growth flags biological wandering, and the slope is a circuit property, not an annoyance.

That slope is the opportunity. Because the model ties the diffusion rate to inhibitory architecture, manipulations that shift inhibitory timescales or footprints should move the wandering rate in predictable directions: development and maturation of inhibition in organoid preparations, or subtype-targeted pharmacology in slice, become measurable as a change in a diffusion coefficient extracted from array data, alongside the qualitative mode signature. The shifting-versus-scaling decomposition is the other half. Near a stability boundary a preparation does not merely drift faster; it changes character, either propagating activity across the array as a traveling wave or rhythmically breathing and collapsing. An array with spatial resolution can tell those two apart from each other and from mere translation; a single electrode or any spatially aggregated metric cannot. This is a concrete, mechanistic argument for spatially resolved long-term recording as opposed to reduced channel counts.

The threat is misattribution in both directions. A preparation quietly approaching a scaling boundary can lose its persistent pattern entirely, which on a long recording looks exactly like the tissue dying or the interface failing; re-referencing or discarding it destroys the most interesting event in the dish. Conversely, a rising wandering slope across days may be maturation of inhibitory circuitry rather than degradation, and treating it as drift to be corrected, or as reason to discard the preparation, throws away the signal. The model gives both failure modes a fingerprint; the instrumentation side has to record with enough spatial and temporal resolution to apply it, which argues for keeping high-density spatial sampling in chronic protocols even when only aggregate readouts are reported.

The bottom line

Established here, within the model: persistent-bump stability is organised by separate shifting and scaling modes; inhibitory subtype thresholds, timescales, footprints, and even weak interneuron-to-interneuron connections reshape the stable region; and stable bumps wander diffusively at a rate that broader SST connectivity slows, with the analytic prediction matching simulation. Established nowhere yet: any of this in tissue. What would confirm it is an array recording in which a sustained activity centroid shows linear variance growth whose slope responds to inhibitory manipulation in the predicted direction, with mode changes near boundaries classifiable as translation or width oscillation. What would break it is tissue whose slow non-stationarity fails the linear-variance signature under conditions where instrument noise is demonstrably negligible. Until then, treat the wandering rate as a hypothesis with a precise, cheap test, and treat unexplained slow drift on your array as data, not dirt.