More electrodes than channels: adaptive routing on a high-density array

What the work claims

This is a method paper with an offline benchmark and a single online demonstration. Its central claim is that under a fixed channel budget an uncertainty-aware Bayesian policy captures more of the available spiking than a static scan-and-commit assignment, because the set of informative electrodes drifts over a long recording.¹ On nine 34 hour recordings the 100 most active electrodes turned over by 47.8 percent by 34 hours, and the Bayesian policy exceeded static selection by 17.2 percentage points of oracle-captured spikes at the final time point. In one online recording it routed 1,024 of 26,400 electrodes, updated the configuration every 30 minutes, and captured the first synchronized network burst.¹

What makes it worth reading is not the novelty of Thompson sampling, which is a textbook bandit method, but the framing. The authors name the readout bottleneck explicitly and treat electrode allocation as sequential decision-making under uncertainty. That is the right abstraction for a piece of instrumentation whose spatial resolution and whose recording capacity are two different numbers.

How the hardware forces the choice

The platform is a switch-matrix CMOS high-density array (a MaxOne, MaxWell Biosystems).¹ Its 26,400 electrodes are pixels that mostly hold a routing switch and a little memory; the low-noise amplifiers, filters, and analog-to-digital converters are costly in silicon area and power, so they sit outside the sensing region and are shared by whichever electrodes are switched onto them. That asymmetry, many cheap electrodes and a few expensive readout chains, is why at most 1,024 electrodes record at once, a multiplexing ratio near 26 to 1. The recorded band is 300 to 3000 Hz, that is the extracellular spike band; slower local field potentials below roughly 300 Hz are filtered out, so this pipeline is spike-driven by construction and says nothing about slow field dynamics.

The algorithm keeps a Gamma belief over each electrode's firing rate, samples one plausible rate per electrode, routes the highest-sampled subset, updates the beliefs of the electrodes it actually read from the observed spike counts, and then discounts every belief by a factor of 0.7 each 30 minute window so old evidence decays. Geometric discounting at 0.7 gives an effective memory of about one divided by one minus 0.7, roughly three windows or near 1.7 hours. In other words the method is deliberately built to forget the past couple of hours and follow slow drift, which is worth keeping in mind against the paper's own observation that burst recruitment also varies over seconds to minutes.

The strongest case for it

The constraint is real and unglamorous, and naming it is a service. A full-array survey on this platform costs 26 configurations of 60 seconds each, about 26 minutes, during which transient population events go unseen, so you cannot simply rescan often to refresh a static map. Given that, a rule that reallocates the budget from partial observations, keeps high-yield sites, and still explores is a sensible answer, and the uncertainty-weighted exploration of Thompson sampling beat both blind random exploration and a simpler epsilon-greedy variant. The authors also give a fair reason to use spike count as a first objective: for spatially bounded activity, capturing a larger fraction of events tightens the achievable error on a center-of-activity estimate, so the yield objective is at least aligned with preserving a spatial summary.

Where a skeptic should push

The load-bearing weakness is that the benchmark is self-refereed. The reward is captured spike count, the oracle is defined as the maximum achievable spike capture, and the score is the fraction of that maximum captured. Beating epsilon-greedy and static toward that oracle shows the policy maximizes the very proxy it was built to maximize; it does not show better scientific yield. There is no independent utility metric here, no count of distinct neurons recovered, no downstream decoding or connectivity accuracy. So 17.2 percentage points is a proxy-optimization margin, not evidence of better science.

The contest is also run on favorable terms. The offline evaluation selects 100 of 529 candidate electrodes, about 19 percent, on a dense 23 by 23 patch that was pre-recorded in full so that the oracle can even be computed. The deployed online problem routes 1,024 of 26,400, about 3.9 percent, with the unselected electrodes genuinely unobserved. The headline margin is therefore demonstrated at a coverage fraction roughly five times more generous than the real task, and the online run is a single qualitative recording. Generalization to the sparse regime is asserted, not shown.

A further limit is structural rather than statistical. Time-shared routing yields sequential coverage, never a synchronous full-array snapshot. Two electrodes that were never co-routed in the same window can never yield a zero-lag cross-correlation, so adaptive routing improves how much you eventually see, not what you can measure at a single instant. Any analysis that needs simultaneity, such as fine synchrony or propagation across distant sites, is bounded by the 1,024 electrodes that happen to share a window, not by the cleverness of the allocator.

Finally, the reward biases the data it collects. Because spike count drives selection, the recorded dataset becomes a function of the policy, and the bias is a rich-get-richer loop: the highest-firing unit pulls selections toward itself, and because a strong soma's extracellular footprint can extend across neighboring electrodes at a 17.5 micron pitch,² summing spike counts can reward redundant coverage of one active source rather than spreading the budget to capture new units. The authors acknowledge both the over-sampling of active regions and the independence assumption. One practical note on the acquisition chain: each time an electrode is switched onto a shared amplifier, the alternating-current coupling and the electrode-electrolyte interface must re-settle before the trace is clean. That is negligible against a 30 minute window, but it sets a floor on how fast re-routing can be useful, which is another reason this method tracks hours-scale drift rather than fast burst-to-burst variation.

Routing as the real array bottleneck

For anyone specifying array hardware, the lesson is that the electrode count on a datasheet is not the number that governs what you can record; the ratio of electrodes to simultaneous low-noise readout chains is. This platform's 26,400 electrodes are a spatial-resolution claim, its 1,024 channels are the acquisition claim, and the two are not the same figure. A policy that reallocates the smaller number is genuine value: it extends effective coverage over a long culture without adding a single amplifier, which is the area-hungry and power-hungry part of the chain. That is the real opportunity, and it is a software opportunity resting on fixed silicon.

The non-obvious implication is that adaptive routing quietly changes the provenance of every dataset it produces. Once the instrument decides what to record based on what it has already recorded, the recording is no longer an unbiased sample of the tissue; it is a closed loop whose bias is activity-dependent. For an organoid program that wants to compare cultures, track development, or screen compounds, that is a threat: two recordings made under two different adaptive trajectories are not directly comparable, and a shift in a network metric may reflect the sampler rather than the biology. The conservative fix is not to abandon adaptivity but to log the full selection trajectory as first-class metadata and to validate any network claim against a fixed-electrode control.

The obsolescence angle cuts the other way. Active-pixel and very-large-scale CMOS arrays that read every electrode, including the field-potential imaging at 236,880 electrodes the paper itself cites, remove the selection problem at the front end and push the bottleneck downstream into data volume and analysis. If that architecture wins, the elegant routing policy becomes unnecessary at the electrode and reappears as a data-reduction policy in the pipeline. Either way the constraint does not vanish; it moves.

For organoids specifically, mind the gap between this preparation and the title's subject. The work is two-dimensional dissociated rat cortex on a planar array. A brain organoid is three-dimensional, and a planar array contacts only the cells resting against the substrate. Adaptive routing can rescue an informative site that has drifted to a neighboring electrode; it cannot route to a neuron in the organoid interior that never touches a contact. For three-dimensional tissue the unobserved-electrode problem becomes an unobserved-interior problem, which is an electrode-geometry problem, not a channel-allocation one. This method addresses the readout bottleneck and leaves the access bottleneck untouched.

The bottom line

Established: on two-dimensional rat cultures, evaluated offline against its own spike-capture oracle at a 19 percent selection ratio, a discounted Thompson-sampling policy captures about 17 percentage points more oracle spiking than a static assignment, and it ran online in one recording. Not yet shown: that the margin survives at the deployed 3.9 percent routing ratio, that it improves any metric other than the one it optimizes, and that adaptively sampled data support unbiased network inference. What would confirm it is an independent utility endpoint, such as distinct units recovered or decoding accuracy, measured online across many recordings, together with evidence that connectivity estimates from adaptive sampling match a fixed-electrode reference. What would break it is a demonstration that the sampling loop distorts the very population statistics these long cultures are recorded to study.