Ensembles of tiny spiking networks, evolved as a team

What the work claims

Rodriquez and Ghawaly at Louisiana State University claim that spiking neural networks too small to solve a task individually can be evolved as a cooperative ensemble, provided each network is selected for what it adds to the group rather than how it performs alone. The credit-assignment rule comes from cooperative game theory and the difference evaluation functions of multiagent systems: a network's fitness is its marginal contribution to ensemble performance, averaged over the groups it joins during evolution. Networks that reliably improve the group are rewarded; redundant copies are penalized; complementary specialists emerge.

This is a primary methods paper, peer-reviewed at ICONS 2026, and its claims are statistical rather than anecdotal: across classification, regression, and control tasks run under the hardware constraints of the Caspian neuromorphic platform, co-evolved ensembles beat both single-network evolution and post-hoc ensembles built by combining independently evolved networks, with one-sided Wilcoxon signed-rank tests reporting significance on every dataset and family. On the CartPole control task, where standard evolution largely fails to discover effective policies, the authors report a qualitative transition to near-optimal performance under co-evolution.¹

How it works

The underlying trainer is EONS, the Evolutionary Optimization for Neuromorphic Systems algorithm in the TENNLab framework, which evolves both topology and parameters of spiking networks without gradients. The scaling argument in the paper is simple and worth taking seriously: for a network of n neurons there are on the order of n-squared potential directed edges, and with hardware-constrained parameter quantization per edge, the searchable space grows super-exponentially in n. Small networks are searchable; large ones are not, within any practical budget. An ensemble of small networks has a combined space that is large, but each member's space stays searchable, and if the members search mostly independently, coupled only through the group fitness signal, mutations in each sub-network explore an exponentially larger fraction of the relevant space than the same mutation would in one monolithic network.

The fitness rule is the contribution. For ensembles of three (chosen because three networks is what typically fits on a Caspian chip), each candidate network's score is the average improvement in group performance over the groups it participates in, relative to that group without it. Because evaluating every partner combination in every generation is cheap for classification and regression but prohibitive for control, the control experiments use a two-phase schedule: an exhaustive first generation followed by a history-biased sampling of partner combinations. The ensembles are not decorated after the fact: group formation is part of evolution, so selection pressure favors networks that are good partners, not networks that are good soloists.

The evidence base is respectable for the regime: classification on Wine, Iris, and Breast Cancer with 30 runs per method per dataset and paired seeds; regression on California Housing, Diabetes, Concrete, and Superconductivity with 75 runs each; and CartPole-v1, with a 500-timestep horizon, 50 runs, and 100 held-out test episodes per trained system. Wilcoxon p-values against both baselines are significant everywhere (for example, 9.15e-06 for co-evolved versus single networks on Wine, and 7.82e-07 and 5.11e-07 for CartPole against the single and post-evolved baselines respectively). Post-hoc ensembling of independently evolved networks, by contrast, yields modest and inconsistent gains, and in control it fails outright: averaging the action outputs of networks that were never trained to be compatible produces policies that do not compose.¹

Where a skeptic should push

The load-bearing assumption is that the ensemble decomposition is solving the intended problem and not relocating it. Marginal-contribution fitness requires evaluating candidate networks across many partner combinations per generation. For classification and regression that is affordable; for control the authors had to switch to a sampled two-phase schedule, and even so the evaluation budget dwarfs what a single-network run spends. The claim that co-evolution rescues tasks where evolution fails needs the caveat that it spends considerably more evaluations to do so. Whether the compute trade is favorable depends on the cost of evaluation, which on silicon is cheap and on living tissue is anything but.

Second, everything here is small. The networks are hardware-constrained toys by machine learning standards, the tasks are tabular benchmarks and one classical control environment, and the ensemble size of three is a hardware quota, not a scaling law. The authors say this plainly in their limitations: they do not claim co-evolution scales to arbitrarily large networks or frontier tasks, and they note that tasks requiring tightly coupled dynamics across many neurons may favor monolithic networks despite the harder search. The super-exponential search-space argument cuts both ways; decomposition buys reach at the price of expressiveness, because an ensemble of small modules can only represent functions that decompose accordingly.

Third, the strongest result, the CartPole transition, comes with higher variance across runs than the baselines, which the authors attribute to the sensitivity of reinforcement learning to small policy differences. Near-optimal median behavior with a wide distribution is still a real improvement over consistent failure, but it is not the same as reliability.

Training organoid collectives without gradients

The non-obvious implication is that this paper describes, almost by accident, one of the few training paradigms that maps onto biological tissue without distortion. Backpropagation is off the table for organoids: no differentiable simulator of a dish, no way to push analytic gradients through spiking living tissue, and closed-loop methods that mimic error signals are young and fragile. But evolution does not need gradients, and marginal-contribution fitness does not need a global model of the substrate. It needs only the ability to form groups, measure group performance, and attribute improvement. A multi-organoid system hooked to a common task interface, where stimulation patterns route inputs and MEA readouts score outputs, satisfies all three. The paper's central finding, that training-time interaction beats post-hoc combination because independently trained members do not compose, will apply with force to organoids: two dishes trained separately and averaged at the electrode will fail for exactly the reason the post-evolved CartPole ensembles failed, their policies are not compatible at every timestep.

The opportunity, then, is a concrete recipe for scaling organoid computing without scaling any single tissue mass: keep each organoid small, searchable, and within its stable regime, and evolve a team against a shared fitness with per-member marginal contribution. That is a plausible answer to the field's growth problem, organoids that get larger get noisier, slower, and harder to keep alive, while a collective of small ones sidesteps several of those failure modes at once.

The threat is the evaluation cost. On silicon, one extra partner evaluation is a forward pass; on tissue it is hours of culture time per configuration, and the combinatorics of partner groups multiply that. A framework that rescues tasks by spending evaluations is a poor fit for the slowest evaluator in computing. There is also a subtler risk: marginal-contribution selection pressures each member toward specialization, and a team of hyperspecialized organoids may be exquisitely fragile to drift, contamination, or the death of one member, the wetware version of the paper's higher CartPole variance. And the dual-use framing deserves a sentence: the same collective-training machinery applies to any population of biological signal processors, not just stem-cell-derived neural tissue, so governance of multi-tissue systems should assume ensemble training is on the menu.

The bottom line

Established: under Caspian-scale hardware constraints, co-evolving small spiking ensembles with marginal-contribution fitness significantly outperforms single-network evolution and post-hoc ensembling across classification, regression, and control, with the decisive advantage appearing where single networks fail. Established with caveats: the gains cost many more evaluations, the networks are small, and the best control result carries high run-to-run variance. For organoid intelligence, the paper's real contribution is a gradient-free, model-free credit-assignment scheme whose only demands are group formation, group scoring, and attribution, all of which a closed-loop multi-organoid rig can in principle supply. What would confirm the transfer is a demonstration on two or three coupled organoids that co-evolved specialization outperforms independently trained and linearly combined tissue; what would break the idea is evidence that the evaluation budget grows faster than wetware culture time allows, or that organoid drift destroys the partner compatibility that the whole scheme depends on.