Adult brains keep learning by rewiring, not just retuning

What the work claims

This is a review and research-roadmap paper, not a primary experiment, and it should be weighted as such: every claim in it is assembled from the literature and filtered through the authors' judgment. Its central claim is that structural plasticity, meaning the creation and removal of synaptic connections rather than the adjustment of their weights, is a distinct and underappreciated learning mechanism in adult neural networks, operating on timescales of hours to days and beyond, while the electrical activity it serves evolves on milliseconds.¹

The review is organized around three questions: what regulates structural plasticity, what it is computationally good for, and what understanding it would enable clinically and technologically. On regulation, the picture that emerges is activity dependent in a specific homeostatic direction: neurons that lose input, through retinal lesions, monocular deprivation, whisker trimming or denervation, grow new spines to compensate, while chronically hyperactive networks, as during epileptic seizures, shed them. On function, the authors argue structural plasticity plausibly underlies stable memory despite continuous synapse turnover, the phenomenon of representational drift, and lesion recovery. On application, they point to stroke therapy design, where homeostatic rewiring models already make testable predictions, and to neuromorphic hardware, where synapses are, as in biology, a limited and energetically costly resource.

How it works

The review carefully separates three things that are easy to conflate. Synaptic plasticity adjusts the strength of existing connections through mechanisms such as long term potentiation, long term depression and spike timing dependent plasticity. Structural synaptic plasticity remodels the morphology of individual synapses, where spine expansion and shrinkage track weight change so closely that size and synaptic strength are effectively coupled. Structural plasticity proper changes the wiring diagram itself: synapses appear and disappear, and the network topology becomes a dynamic variable. The authors are explicit that this is a slower axis than synaptic plasticity, and that the difference in timescale implies a difference in function rather than mere sloppiness.

The mechanistic spine of the review is the homeostatic structural plasticity model line associated with Butz, van Ooyen, Wörgötter and colleagues, in which each neuron tracks its own long-run average firing activity and adjusts synapse formation and elimination to hold that activity near a set-point. Low activity promotes the growth of new spines; high activity promotes their removal. This simple rule has an impressive reach: the same model class reproduces coarse features of cortical layer-to-layer connectivity after deafferentation, mimics aspects of recovery after focal stroke, and, in the variant implemented in the NEST simulator, can run continuously inside a large network simulation without pausing state propagation. Analogous rewiring machinery exists on dedicated neuromorphic platforms: SpiNNaker demonstrated structural plasticity on both its first and second generation systems, and BrainScaleS-2 handles rewiring in an embedded digital processor sitting beside its analog neuron cores.

One result the review highlights as a genuinely cross-scale phenomenon is the interaction of slow structural change with fast spike timing plasticity. Work combining the two shows that their interaction, rather than either alone, can explain prominent network reorganization effects seen with coordinated reset stimulation, a stimulation protocol originally developed for deep brain stimulation and now being extended to non-invasive vibrotactile treatment in Parkinson's disease. The general principle the authors draw is that fast and slow plasticity are coupled control loops: synaptic plasticity shapes activity on the fly, and structural plasticity then reshapes the hardware that fast plasticity runs on.

Where a skeptic should push

The single most load-bearing assumption is that structural plasticity is doing computational work at all, as opposed to being maintenance. The evidence is real but mostly observational: longitudinal two-photon imaging shows spines appearing and disappearing over days, and the phenomena are experience dependent and lesion responsive. What is thin is causal closure. The review itself lists as open problems exactly the questions a skeptic would ask: how rewiring is controlled at the molecular level, what the network-level consequences of spatially heterogeneous rewiring are, and whether structural plasticity alone can account for adult learning or merely one component of it.

Second, the models are validated on statistics, not predictions. The Butz and van Ooyen class of rules reproduces observed distributions of connectivity and coarse cortical backbones, but reproducing a distribution after the fact is a much weaker standard than predicting the rewiring of a particular network under a particular perturbation. The review is honest that single-time-point connectomics, however large, cannot see any of this: a static wiring diagram is one frame of a film, and the field's flagship connectome datasets are all single frames.

Third, quantitative discipline is scarce in this literature, and the review inherits that. It contains almost no numbers: formation and elimination rates are not tabulated, dose-response relationships between activity and rewiring are not summarized, and the memory-stability claims rest on modeling studies of equifunctional networks, where memory is argued to survive turnover because connectivity patterns rather than individual synapses carry the trace. That is a plausible hypothesis with elegant demonstrations in simplified networks, not an established fact about adult cortex. Finally, forty-author reviews smooth disagreements; the strong version of the structural plasticity program is stated here more confidently than the primary literature uniformly supports.

Rewiring is the training organoids already own

For organoid intelligence the non-obvious implication cuts against the field's reflex. Training protocols for living neural tissue overwhelmingly target fast plasticity: spike timing dependent protocols, patterned stimulation, reinforcement-style dopaminergic or optogenetic feedback, all designed to work within an experimental session or a few days. This review says the tissue is running a second training loop on its own schedule, one that no current protocol engages deliberately, and that the loop is regulated by precisely the variable a closed-loop rig controls: long-run average activity.

The opportunity is a coarse but honest write mechanism. Homeostatic structural plasticity predicts, at the level of network topology, which direction chronic stimulation pushes a culture: sustained underactivity grows new connections, sustained overactivity prunes them. That is a one-bit, slow, system-level write channel into living tissue, available without genetic manipulation and without touching individual synapses. It is also the only training channel whose timescale matches the tissue's own self-organization. A serious closed-loop organoid program should treat weeks-long activity shaping as a first-class protocol alongside session-scale plasticity, and should measure topology change, not just firing rates, as its training readout.

The threat is the same phenomenon read from the experimenter's side: the substrate you are training is training itself. Topology turnover means every weight readout has a shelf life, that a decoder fit on day three of a culture is describing a network that no longer exists by day ten, and that representational drift, which the review frames as a natural consequence of turnover, sets a hard ceiling on how static any long-lived organoid system can be. There is no fixed circuit to reverse engineer. The equifunctionality result sharpens this: if memory in these systems lives in statistical connectivity patterns rather than in identifiable synapses, then interpretability approaches borrowed from deep learning, which assume stable parameters, are structurally the wrong tool.

A final observation belongs in the obsolescence ledger. Neuromorphic hardware is busily implementing structural plasticity in the loop, on SpiNNaker and BrainScaleS-2, motivated by the same insight this review develops for biology: synapses are a limited, expensive resource and topology should be managed, not fixed. If both the living substrate and its silicon competitors converge on reconfigurable-topology computing, the differentiator for tissue will not be that it rewires, which silicon now also does on demand, but the biochemical richness of what a synapse is when it forms. That is a much narrower claim than the field usually makes.

The bottom line

Established: adult neural networks continuously form and eliminate synapses on hour-to-day timescales; the process is activity dependent in a homeostatic direction; it responds to lesions and deprivation with compensatory rewiring; and homeostatic rewiring models reproduce a real range of experimental observations in simulation. Established also, on dedicated platforms: structural plasticity can run in the loop in both software and neuromorphic hardware. Hypothesis, not established: that structural plasticity is the principal substrate of adult learning and stable memory, that its rules are known well enough to exploit predictively, and that coordinated reset style stimulation works through the structural-fast plasticity interaction the review proposes. What would confirm the strong claim is a closed-loop experiment that prescribes a target topology change via chronic activity shaping and observes the predicted rewiring on schedule; what would break it is evidence that adult rewiring is dominated by factors other than activity set-points, which several of the review's own open problems concede is possible.