The map of one mouse's cortex makes better learners

What the work claims

This is a primary computational result with a clean factorial design. The MICrONS dataset provides electron-microscopy reconstructions of more than 200,000 cells and 523 million synapses in mouse visual cortex, co-registered with two-photon calcium imaging of roughly 75,000 neurons across four visual areas and layers 2 to 5; for about 12,000 excitatory neurons both structure and function are available for the same cells. The authors build recurrent networks whose hidden layer matches a sampled cortical field (312 units), then vary three separable priors: function-derived weight initialization, real neuronal coordinates as the spatial embedding, and communicability-aware regularization that discourages long or inefficient communication paths. Each model is trained with Adam on three small cognitive tasks, one-choice inference, perceptual decision-making, and Go/NoGo, across 20 independent runs of 10 epochs.¹

The central claim is that cortical organization is a usable inductive bias, not just a description. The fully grounded variants reach mean accuracies of 0.917, 0.865, and 0.948 across the three tasks, and up to 0.985 on the first task, while several minimally constrained variants sit at chance (0.25 on the four-choice task, 0.50 on the two-choice tasks). Under a positive-only constraint, mimicking an excitatory-only regime, every randomly initialized model collapses to chance, while functionally initialized models retain accuracies up to 1.000, 0.838, and 1.000. The constrained networks also converge to topologically structured weight organizations, low entropy, more modular, more small-world, rather than the random-like configurations unconstrained training finds.

How it works

An inductive bias is any constraint that narrows the space of solutions a learning algorithm searches. Here the biases come from measured biology rather than design intuition. The functional initialization converts pairwise activity correlations among the 12,000 neurons, computed from deconvolved calcium traces, into the initial recurrent weight matrix, so learning starts from the statistical structure the real circuit actually exhibits rather than from random weights. The spatial embedding places each hidden unit at its neuron's true coordinates, so distance-dependent wiring costs are real. The communicability regularizer, built on a walk-sum measure of how directly two neurons can exchange signals, penalizes connections that anatomically would route around the cortex inefficiently. Because the three priors are applied factorially, with Kruskal-Wallis omnibus tests and Holm-corrected pairwise comparisons across the 20 runs, their contributions can be separated rather than lumped into a single "biology helps" effect.

The ablation yields a clear hierarchy. Functional initialization provides the largest and most general gain. Real spatial coordinates add a robust secondary benefit, visible even in models that lack biological initialization, which establishes that cortical geometry itself, not just the wiring derived from it, is a computational prior. Communicability contributes smaller, task-dependent refinements. Two control experiments sharpen the interpretation. Permuting the assignment of biological weight values to neuron positions (W!) leaves performance statistically indistinguishable from the faithful mapping (W*), so the gain rides on the statistics of the biological weight distribution rather than on neuron-to-neuron identity. Yet resampling the same distribution (ECDF resampling) while preserving its marginals degrades performance, so not every distribution-matched matrix is equivalent; the specific correlational pattern matters. The positive-only experiment is the most striking: with all recurrent weights constrained to be non-negative, random initialization can no longer find any solution, while biologically initialized networks still train, because the prior has already pushed them into a favorable region of an otherwise hostile optimization landscape.

Where a skeptic should push

The most load-bearing assumption is that gains measured on three toy tasks generalize to computation worth doing. One-choice inference, perceptual decision-making, and Go/NoGo are decades-old probes with chance levels of 0.25 and 0.50; a 312-unit recurrent net either clears them or does not, which is a coarse instrument for the fine structure of an inductive bias. The tasks are also trained by backpropagation, so the result concerns initialization, not biological learning; nothing here shows that cortical priors would help a network trained by local, biologically plausible plasticity. The data come from one animal, one cortical system, and excitatory neurons only; the authors note the omission of inhibitory cell types, which are central to cortical dynamics, and the positive-only regime is itself an excitatory-only caricature. Finally, the positive-only collapse deserves a second look: it is as much a statement about how badly signed networks train under a naive optimizer as it is about the power of biological initialization. None of this overturns the result, but it bounds the claim: demonstrated on small tasks with full gradients, plausible but unproven at scale, in spiking tissue, or under local learning rules.

The cortical prior organoids never had

The non-obvious implication for organoid intelligence is that this paper quietly reranks what a dish is missing. The standard worry is anatomical: organoids lack cortical lamination, canonical wiring motifs, and the geometry of a real cortex. But in this decomposition anatomy is not the dominant prior. The largest gain comes from the activity-derived functional matrix, the correlational structure of what the population does, not from where the cells sit. That is something an organoid on a microelectrode array already yields: spike recordings from a culture give you exactly the raw material for a functional correlation matrix, without a single electron-micrograph. The practical blueprint is a calibration step that OI practice currently skips. Before training a readout on a culture, measure its functional correlation structure and initialize or constrain the downstream recurrent model with it, the way this paper initializes with cortical correlations; and when comparing dishes, compare their functional statistics rather than their firing rates alone. The hierarchy in the ablation tells you what to measure first.¹

The threat cuts in the opposite direction. Geometry is secondary but real: real spatial embedding added reliable gains on every task, and organoids do not have cortical geometry to offer. More pointed is the positive-only result. An excitatory-only regime is the caricature of organoid preparations with weak or absent inhibitory integration, and the finding is that random substrates in that regime are effectively untrainable on even trivial tasks; you only escape via the right initial statistics. A dish whose functional structure you have not measured may be precisely the randomly initialized, sign-constrained case that collapses. The deeper risk is misreading the mechanism: if the field treats "biology helps" as a slogan, it will miss that the help here comes from measured statistics injected before learning, which means the advantage belongs to whoever characterizes their substrate, not to biology as such. Tissue without metrology is just another bad initialization.

The bottom line

Established: co-registered cortical structure and function, used as initialization and constraint, materially improve recurrent learning on small decision tasks, and the functional component of the prior contributes the largest share; the effect is statistically decomposed, replicated across 20 runs, and robust to permuting neuron identity. Open: whether the hierarchy survives richer tasks, spiking dynamics, local plasticity rules, and non-cortical or disorganized tissue. For biological computing the actionable idea is substrate-aware initialization: characterize a culture's functional correlation structure from its own activity and use it as a training prior, and treat that characterization as a quality-control metric in its own right. What would confirm it: an organoid experiment showing that readouts initialized from a culture's measured functional statistics learn faster and generalize better than ones initialized randomly. What would break it: evidence that culture-to-culture variability swamps the functional structure, or that the gains evaporate once learning is restricted to local, biologically deliverable rules.