An external modulator reconfigures a fixed analog reservoir, and the controls matter more than the headline number
A 60-neuron recurrent network on a DYNAP-SE mixed-signal chip is wired once and never retrained. An off-chip neural network reads its input and activity and rewrites two population-level synaptic gains every 40 milliseconds, lifting binary spoken-digit accuracy from 71.4% to 88.6%. The gain survives controls that rule out retraining, static tuning, and input-independent tricks.
Source: Input-dependent Neuromodulation Enhances Mixed-Signal Neuromorphic Hardware, arXiv:2610.41265, 30 October 2026. Primary source. Read: full PDF, verified inline against every number quoted here.
What the work claims
Mehta, AlKilany, Goodman, Indiveri and Payvand demonstrate input-dependent, within-trial neuromodulation of a spiking recurrent network running on physical mixed-signal neuromorphic silicon. The on-chip recurrent network is a fixed reservoir: its input and recurrent weights are random and frozen. A separate modulator network, running off chip, watches the 96-channel input and the activity of all 60 on-chip neurons, and every 40 ms updates three gains: the excitatory synaptic gain onto the excitatory population, the inhibitory synaptic gain onto that population, and the gain of the off-chip readout. Across five independently instantiated physical reservoirs, five-versus-nine spoken-digit classification rises from 71.4% to 88.6%, a mean improvement of +17.3 percentage points with a standard deviation of 10.4 across chips (two-sided paired t-test, p = 0.021; Benjamini-Hochberg corrected q = 0.082)1. The improvement is positive on every one of the five chips.
This is a primary result paper, and the mechanism claim is narrow and concrete: a small number of dynamically controlled, population-level gains can increase the computational capability of an otherwise fixed and heterogeneous analog substrate. Nothing is reweighted; the anatomical wiring never changes during a trial or between trials.
How it works
The chip is the DYNAP-SE, fabricated in 0.18 micrometer CMOS, with 1024 adaptive exponential integrate-and-fire neurons across four cores. Its synapses are Differential Pair Integrator log-domain circuits, and the key circuit fact is that a DPI synapse separates the synaptic weight from a distinct gain-modulation current. In the circuit equation the output synaptic current scales as weight divided by gain current, so a single shared bias per core multiplies the effective strength of every synapse of that type, on a timescale limited only by how fast the bias can be rewritten. That shared analog knob is what the authors exploit: it is, in effect, a hardware implementation of gain modulation, the neuromodulatory principle by which a fixed anatomical circuit moves between dynamical regimes without rewiring1.
The reservoir is deliberately tiny: 48 excitatory and 12 inhibitory neurons on separate cores, which enforces Dale's law and gives the two populations independent shared gains. The task is binary classification of digits five and nine from the Spiking Heidelberg Digits dataset, chosen because a sweep over all binary pairs found five-nine is where small recurrent networks beat chance and beat feed-forward baselines. The 700 input channels are compressed to 96 to fit the 64-synapse fan-in limit, and testing uses the 213 trials of the standard test split, which contains two speakers never heard during training.
Because gradients cannot propagate through the analog dynamics, the modulator is trained chip-in-the-loop with a surrogate: before training, the team records reservoir responses over a 4 by 4 grid of the two on-chip gains and fits a quadratic model of each neuron's firing rate as a function of the log-gains. During training, gradients from the off-chip readout are routed through these fitted response surfaces to estimate how loss depends on each gain. The modulator itself is a small multilayer perceptron with 128 ReLU hidden units and three tanh outputs; it integrates its outputs into the log-gain each 40 ms window, and the result is soft-clipped to a band of baseline gain divided or multiplied by 1.3, the range the DPI bias circuits can tolerate1.
Three controls pin down the causal story. Control A retrains the readout with a fresh optimizer for the same number of epochs without modulation and reaches 83.9%, 4.7 points below the modulated system, so extra readout training does not explain the gain. Control B replaces the dynamic gains with a single constant setting equal to their validation mean, which costs 18.8 points, so the system is not merely finding a better static operating point. Control C plays each trial the gain trajectory computed for a different trial, which costs 14.1 points, so the gains must actually depend on the current input. Holding only the off-chip readout gain at its mean costs just 0.4 points, confirming the effect lives in the on-chip synaptic gains1.
Where a skeptic should push
The single most load-bearing assumption is that a two-sided paired t-test over five chips justifies a general claim about analog substrates. It does, barely, and only if you accept the corrected q = 0.082 as evidence rather than the raw p = 0.021. Five physical instantiations is a small sample, and the standard deviation of 10.4 points across chips is large: some of that spread is device mismatch, which the authors note can reach coefficients of variation up to 0.2. The honest reading is that dynamic gain modulation helps on every chip measured so far, and that the effect size is genuinely large, but the statistical support is closer to the boundary than the abstract's confidence suggests.
Second, scope. One binary task, two modulated on-chip gains, a 60-neuron reservoir. The abstract is careful to say the work considers exactly this regime and nothing more; the modulator and readout are host software, so the system is not the fully hardware-native neuromorphic processor the conclusion gestures toward. Third, the modulator observes the reservoir's own activity, which raises a subtle circularity: part of what it modulates is its own input statistics. The authors do not address whether the modulator is partly compensating for per-chip idiosyncrasies, that is, learning the mismatch rather than the computation. Demonstrated on one chip family, on one task, with surrogate gradients fitted from a coarse grid: the mechanism is proven, the generality is asserted.
Neuromodulation as an organoid control channel
For organoid intelligence the interesting move is not the accuracy. It is the architecture: a fixed, imperfect, slowly-changing biological or physical substrate whose effective computation is steered by a fast external loop that adjusts a handful of global parameters. That is precisely the operating regime every cortical organoid computing setup already lives in. You cannot rewire a culture on demand; you cannot retrain its synaptic matrix; you can, however, modulate it. The DYNAP-SE result gives this a worked existence proof on silicon and, more usefully, a price list: three population-level gains, updated at 25 Hz, bought 17 points of accuracy on a substrate with up to 20 percent device variability. Tissue has far more modulation channels than this chip does, neurotransmitter baths, opsin-expression patterns, tonic field stimulation, temperature, but no clean separation between weight and gain, and no datasheet for what any knob does.
The opportunity is a concrete experimental blueprint. Pair an organoid on a multielectrode array with a modulator network trained exactly as here: record the culture's response to a coarse grid of one or two global stimulation or pharmacological settings, fit per-neuron response surfaces, then train the modulator through the surrogate while the culture computes. This reframes the central organoid problem, what a given culture can compute, into a problem of regime selection, where a fixed culture with several accessible dynamical regimes may outperform a nominally better culture locked into one. The silicon result also quietly undercuts a common organoid pitch: if dynamic gating is what makes small fixed substrates useful, much of the perceived computational advantage of living tissue could in principle be matched by neuromodulated silicon at a fraction of the engineering risk.
The threat is auditability. The surrogate-gradient method works by fitting a model of the substrate, and a fitted model of a culture will happily absorb batch effects: passage number, plating density, recording day, electrode drift. A modulator trained through that surrogate can learn to exploit those confounds, and every one of the paper's three controls has a wetware analog that almost nobody runs. Playing a trial the gain trajectory from another trial, their Control C, is the single most transferable experimental design in this paper: it is the difference between a closed-loop system that reads the tissue and one that replays a script.
The bottom line
Established: on the DYNAP-SE, dynamic, input-dependent modulation of two shared synaptic gains reliably improves a fixed analog reservoir, the effect is causal by the paper's own controls, and it does not reduce to readout retraining or static tuning. Not yet established: generality across tasks, gain counts, and substrate scales, and the q = 0.082 corrected value means the statistical case should be read as suggestive rather than decisive. What would confirm it is the same experiment on a harder task with more modulated populations and more chips. What would break it is a demonstration that surrogate-trained modulators mostly compensate for mismatch rather than expand computation, which the current experiments cannot rule out. For organoid intelligence, the durable export is the control design, not the number.
Frequently asked questions
What is a DYNAP-SE chip and why does its synapse circuit matter here?
DYNAP-SE is a mixed-signal neuromorphic processor with analog subthreshold neuron and synapse circuits. Its Differential Pair Integrator synapses expose a gain-modulation bias separate from the weight, which is the circuit primitive that makes within-trial gain rewriting possible.
Why could the authors not just train the on-chip network directly?
Gradients cannot be backpropagated through the chip's analog dynamics, so the recurrent weights stay fixed and random. The team instead fitted quadratic response surfaces of each neuron's firing rate to the two gain parameters, and trained the modulator through those surrogates.
How large is the accuracy improvement and how confident should we be?
Mean improvement is +17.3 percentage points across five physical reservoirs, from 71.4% to 88.6%, with p = 0.021 uncorrected and q = 0.082 after multiple-comparison correction, so the effect is large but the statistical support is near conventional boundaries.
Which control was the most decisive?
Control C, which replayed each trial the gain trajectory computed for a different trial, costing 14.1 points of accuracy. It shows the modulation must be computed from the current input, not merely present.
Does this mean small fixed organoids could compute more than we assume?
It suggests computational capability of a fixed substrate is partly a property of how it is steered, not only of its size or wiring, which is a real and testable hypothesis for cultured neural tissue coupled to a fast modulation loop.
References
- Mehta A, AlKilany A, Goodman DFM, Indiveri G, Payvand M. Input-dependent Neuromodulation Enhances Mixed-Signal Neuromorphic Hardware. arXiv:2610.41265. 2026. https://arxiv.org/abs/2610.41265. Accessed 2026-11-02.