Keeping a decoder alive while the tissue underneath it drifts

What the work claims

Intracortical brain-computer interfaces (iBCIs) decode intended movement from electrodes implanted in cortex. Their chronic weakness is not accuracy on day one but decay afterwards: the recorded population shifts from session to session as electrodes micromove, tissue responds, and units appear and vanish, so a decoder trained on one day gradually stops generalizing.¹ The paper proposes Membrane Potential Alignment (MPA), a way to keep a pretrained spiking decoder working on later days without collecting any new labelled movements.

The reported result is concrete. On a single non-human primate performing self-paced grid reaching, evaluated on ten target sessions drawn from a month-plus recording span (the benchmarked targets themselves run out to about three weeks), MPA raised the mean velocity-decoding coefficient of determination (R squared, the fraction of movement variance the decoder explains) from 0.09 to 0.33 relative to a non-adapting baseline, improving nine of ten target sessions. It was competitive with NoMAD, a heavier recurrent alignment method that the paper reports held decoding stable for over three months, and it did so while adapting fewer than nine percent of its parameters and running at four millisecond resolution instead of twenty.¹ This is an engineering paper with a measured benchmark, not a theoretical proposal, so it should be weighed on its numbers and its failure cases rather than on ambition.

How the alignment works

The decoder is a spiking neural network of leaky integrate-and-fire units: each unit accumulates input into a membrane potential and emits a spike when it crosses threshold. MPA rests on one observation. When the input statistics drift on a later day, that drift propagates through the network and shows up as a shift in the distribution of internal membrane potentials. So instead of trying to correct the raw spikes or retrain the output, MPA models each hidden layer's membrane potentials as a multivariate Gaussian and minimizes the Kullback-Leibler divergence (a measure of how far one probability distribution sits from another) between the new day's distribution and the original day-zero distribution.¹

Two design choices carry the method. First, only low-rank adapter (LoRA) weights in the first hidden layer are updated while the trained decoder stays frozen, which is why under nine percent of parameters move; the learned input-to-velocity mapping is preserved and only the front end is nudged back into register. Second, the neurons use a soft reset that leaves residual membrane potential after a spike rather than zeroing it, which keeps the membrane distribution continuous and gives the alignment objective a usable gradient. The payoff is data efficiency: the paper reports restoring decoding with as little as 1.88 seconds of source data and 1.28 seconds of alignment data, and in one representative session it lifted R squared from about 0.08 to 0.46 on both velocity axes.¹

Where a skeptic should push

The load-bearing assumption is that a decoder trained on day zero still captures enough of the underlying neural structure that realigning an internal distribution can recover it. The paper is unusually honest that this assumption fails. Both MPA and NoMAD collapse on one session (Day 8) where the shift is too large for label-free alignment, and the authors add explicit compatibility metrics, representational similarity and a measure of how much the low-dimensional neural subspace has rotated, precisely because unsupervised adaptation can otherwise fail silently, producing confident but wrong velocity predictions. That silent-failure mode is the real hazard, and it is worth stating plainly rather than burying.

Three further cautions. The demonstration is a single subject on one reaching task, so the month-long numbers are an existence proof, not a distribution over animals or behaviours. The absolute performance is modest (a mean R squared of 0.33 leaves most of the movement variance unexplained), which reflects a deliberately harder self-paced dataset than the structured center-out task NoMAD was originally tuned on, but it also means the method rescues a mediocre decoder rather than a good one. And matching a Gaussian model of membrane potentials assumes the drift is well described by first and second moments; a reorganization that changes the shape of the distribution, not just its mean and covariance, is exactly the regime where the compatibility metrics say to stop.

Reading a living substrate that will not hold still

Organoid intelligence borrows the iBCI stack wholesale: you stimulate living neural tissue, record its spikes on a microelectrode array, and train a decoder to read intent or computed output from that activity. The unglamorous truth this paper surfaces is that the readout is where much of the fight is, and it is a moving target. A cortical organoid is plausibly far less stationary than an implanted primate cortex, though that is an expectation rather than something this paper shows: it is still developing, rewiring, and changing its firing statistics week to week by construction, so the drift that degrades an iBCI decoder over weeks may be a mild version of what a developing organoid readout faces. An organoid-computing benchmark that trains and tests a decoder within one session can therefore flatter itself; the more honest metric is cross-session, label-free performance, which is the regime MPA was built for.

The non-obvious implication is about where learning is allowed to happen. In MPA the neurons are frozen and only the silicon adapter moves. With a living substrate you have no choice at all: you cannot LoRA-adapt biological neurons, and you rarely have ground-truth labels for what an organoid was supposed to compute. So a large share of the burden of stability has to sit in the decoder, carried by aligning an internal statistic without supervision, with the Day 8 collapse a reminder that decoder-side alignment alone can still fail outright when the shift is large enough. That reframes a common claim about wetware. The romance of organoid computing is that the tissue learns; what this paper suggests is that much of any working system's adaptation is silicon compensating for a substrate that will not stay put. The opportunity is that the recipe is cheap and portable: model an internal distribution, align it with a few seconds of unlabelled data, and gate it with a compatibility check. The threat is the silent-failure mode, sharpened for tissue. An organoid can reorganize past the point where alignment is valid while the decoder keeps emitting confident outputs, and unlike the primate case there is no behavioural ground truth to catch it. A drift detector that halts and flags, rather than one that always adapts, looks less like a nicety than a necessary safeguard for living substrates. One caveat bounds all of this: MPA's mechanism is Gaussian-matching of a spiking decoder's own membrane potentials, so the transfer presupposes the organoid readout is itself a spiking decoder with an internal distribution to realign; a non-spiking organoid decoder would need a different alignment target, not this exact method.

The bottom line

Established result: on one primate over a month, aligning a frozen spiking decoder's membrane-potential distribution with under nine percent trainable weights and a few seconds of unlabelled data recovers most of the decoding lost to session drift, and fails predictably when the shift is too large. Hypothesis, not result: that the same trick transfers to the harder, faster, label-free drift of a developing organoid. What would confirm it is a cross-session organoid study that reports label-free readout stability and, crucially, characterizes the failure boundary. What would break it is drift dominated by distribution-shape changes rather than shifts in mean and covariance, where a Gaussian alignment has nothing to grip. Either way, the paper relocates the real problem in organoid computing from teaching the tissue to reading it reliably as it changes.