Local cycles and interneurons decide what a network can compute

What the work claims

This is a computational study, a systematic mapping of network structure onto computational ability rather than a neuroscience experiment, and its authority comes from exhaustiveness rather than any single dramatic measurement. Talpir and Schneidman train recurrent neural networks to compute Boolean functions, mappings from N binary inputs to one binary output chosen as the task space because it is complete: for N inputs there are exactly 2 to the 2-to-the-N such functions, so a network that can compute all of them can, in this strict sense, compute anything of that input width. For small networks they enumerate everything: all 64 directed networks of three neurons are trained on all 256 three-bit functions, and all 4096 networks of four neurons are trained on all 65536 four-bit functions, 500 epochs of backpropagation through time, ten random initializations each, with success defined as correct output on every input in at least one initialization.¹

The resulting catalogs deliver three findings. First, capacity is wildly uneven: most networks are poor, and most functions are hard, with success concentrated in particular architectures and particular function classes. Second, networks containing local 2-cycles (reciprocally connected neuron pairs) and 3-cycles (directed loops of three neurons) are systematically more capable, and are frequently the minimal architectures that solve a given function: strip the cycle and the function becomes unlearnable at that size. Third, a linear predictor built from just three scalar counts, the number of connections, the number of 2-cycles, and the number of 3-cycles, captures most of the predictive power of the full connectivity matrix, cross-validated across 20 train/test splits.¹

The extension to large networks is where the result bites. Sampling Erdos-Renyi random networks, the authors find that typical networks fail as size grows, in the strong sense that they cannot even approximate a randomly chosen Boolean function. Then comes the surprise: adding a small number of sparsely connected interneurons, neurons that receive no external input and participate only in the recurrent dynamics, dramatically restores the ability to learn, even in very sparse networks that were useless without them.¹

How it works

The catalog construction is the methodological core. For each network-function pair the network is initialized with a binary input pattern and its trained recurrent dynamics must settle on the correct output, a paradigm close in spirit to reservoir-style computation where temporal evolution, not a static lookup, does the work. Two scoring matrices result: a Catalog Matrix recording which network exactly computes which function, and an Approximation Matrix recording the best accuracy achieved, which exposes the gap between nearly computing something and computing it. Quotienting out symmetry, relabeled networks and input-permuted functions, condenses the combinatorial explosion into a structured tree in which cycle-rich architectures sit at the leaves solving the broadest range of functions.¹

The controls are what make the cycle claim credible. To ask whether cycles are just proxies for good information propagation, the authors test directed acyclic graphs with matched connection density, which perform poorly at every size, and then a sharper construction: networks engineered for high reachability in few hops while 3-cycles are explicitly prohibited. These reachability-matched acyclic networks still fail, often worse than ordinary DAGs, showing that propagating input quickly is not the same as retaining and re-processing it, and that recurrence, specifically short closed loops, is doing a distinct computational job. An input-expanding DAG variant that forces broad projection from input neurons partially rescues performance, so input spreading matters, but it does not close the gap to cyclic networks.¹

In the large-network experiments, the interneuron augmentation is deliberately minimal: random networks of ten input neurons plus additional interneurons, 100 architectures sampled per configuration, each trained on 100 random ten-bit functions. Learnability recovers rapidly as interneurons are added, and the effect is largest precisely where the base network was worst, consistent with the cycle statistics of random graphs, which grow abundant in short cycles as size increases: the interneurons appear to be buying recurrent closure, not just extra parameters.

Where a skeptic should push

The most load-bearing assumption is that coverage of the total space of Boolean functions is a meaningful yardstick of computational ability. It is the harshest possible yardstick, and the authors are honest that most functions are intrinsically hard in a complexity-theoretic sense. Real nervous systems never face uniform draws from all Boolean functions; they solve structured, low-dimensional, ecologically shaped problems, for which the capacity ranking across architectures could look quite different. A network class that loses on random functions might dominate on the function family biology actually uses.¹

Second, the catalogs conflate computation with learnability under one specific regime. What is measured is what gradient descent through time finds from random initializations in 500 epochs, not what the dynamics permit in principle: a network recorded as unable to compute a function might reach it under another training rule, another readout, or longer training. That is a real gap between "cannot compute" and "this training procedure did not find the computation," and it cuts both ways, since the training process is itself part of any future organoid-training pipeline. Third, the step from four neurons to ten, or twenty, is still orders of magnitude below biological scale, and the interneuron rescue, while well controlled against acyclicity and reachability, is demonstrated in idealized rate networks rather than spiking or biophysical models. Treat the design principles as strong evidence within the model class, not as theorems about tissue.

Cycles and interneurons set the organoid ceiling

For organoid intelligence, the uncomfortable implication is that the field's quiet default, that more cells means more compute, is exactly the assumption this paper dismantles. Random large networks are computationally empty; capacity is a property of local structure, of whether the wiring contains short recurrent loops, that no current organoid protocol controls. A cortical organoid is grown, not laid out: its effective connectivity emerges from self-organization, varies batch to batch, and nobody selects for cycle abundance. If this paper's logic transfers, most organoid networks may sit in the long tail of low-capacity architectures, and scaling cell count without controlling topology would buy almost nothing: a direct hype-correction for "bigger dish" narratives, and a reason the field's unit of progress should shift from neuron number to connectivity statistics.¹

The hopeful half of the result is the interneuron rescue, because it is the one mechanism biology already implements spontaneously. Cortical organoids generate inhibitory interneuron lineages that integrate into the network, and under this paper's account those cells may be doing far more computational work than their numbers suggest: they close recurrent loops and let sparse, weakly connected tissue retain and re-process signals it would otherwise dissipate. Maturation protocols that enrich inhibitory diversity, or assembloid designs that supply it, plausibly raise a culture's computational ceiling more efficiently than any increase in excitatory cell count. This is a testable, mechanistic reason the excitatory-inhibitory balance work in the organoid literature matters to biocomputing specifically, not only to disease modeling.

The non-obvious opportunity is evaluative. If three scalar counts, connections, 2-cycles, 3-cycles, predict a network's capacity, then a structural assay of a living culture, built from high-density multielectrode cross-correlograms or perturbation-mapping experiments, could estimate its computational ceiling before anyone spends weeks on closed-loop training. You would assay the culture's effective wiring rather than its size, screen for cycle-rich batches, and concentrate training where the substrate can actually pay it off. The threat that comes with it is a substrate lottery with a selection step the ethics literature has not caught up to: if capacity is structural and screening becomes routine, the field will be choosing which living neural tissues are "worth" computing on, which is a governance question as much as an engineering one. And the same structure-function framework applies to synthetic reservoirs, meaning a competitor that can etch cycle-rich topology on demand starts with an advantage no grown substrate can currently match.

The bottom line

Established, with unusual rigor for the model class: exhaustive enumeration shows most small recurrent networks are poor computers, local 2- and 3-cycles strongly enhance and often minimally suffice for computation, three structural statistics predict capacity, typical large random networks fail even at approximation, and adding sparsely connected interneurons rescues them, against controls that rule out mere propagation as the explanation. Not established: that Boolean coverage is the right measure of ability for structured real-world tasks, that the training-free dynamics (rather than BPTT findability) carry the limitation, or that any of this survives translation to spiking, biophysical, or grown networks. What would confirm transfer is a demonstration that cycle abundance in organoid effective connectivity predicts closed-loop task learnability across batches; what would break it is a finding that structured, naturalistic tasks erase the capacity differences the catalogs expose. For organoid intelligence the durable lesson is simple to state and hard to act on: assay and select for structure, not size, and take inhibitory interneuroids seriously as computational hardware.