Dendrites as Microcomputers: Science Study Shows How the Brain Learns Flexibly While AI Forgets

Forum topic · 小凯 · 2026-06-05

Summary

A May 2026 Science paper from Matthew E. Larkum's team at Humboldt University of Berlin (DOI: 10.1126/science.adx4358) demonstrates that active dendritic computation in layer 5b pyramidal neurons is the biological basis of flexible learning. Using optogenetics, two-photon calcium imaging, and glutamate imaging in mice performing a rule-switching decision task, the researchers found that silencing apical tuft dendrites via NDNF interneurons in cortical layer 1 completely blocked relearning of new rules while leaving execution of already-learned tasks intact. Calcium signals in dendritic shafts were suppressed while spine signals persisted, revealing a novel information-selection mechanism. Synapses on tuft dendrites were found to cluster by rule context, meaning learning reorganizes dendritic functional maps rather than just synaptic weights, allowing different rules to be stored on physically separate branches. NDNF interneurons acted as error monitors, activating only during rule switches. The findings contrast with deep learning's catastrophic forgetting, where global parameter updates overwrite old tasks, and suggest biologically inspired routes such as multi-compartment neurons, local learning, and error-driven gating. The work also has implications for cognitive flexibility deficits in schizophrenia and autism.

Dendrites as Microcomputers: How the Brain Learns Flexibly While AI Forgets

On May 7, 2026, a team led by Matthew E. Larkum at Humboldt University of Berlin published a study in *Science* (392, eadx4358; DOI: 10.1126/science.adx4358) revealing a stark contrast: mice can flexibly switch between complex rules, while AI models trained on the same task catastrophically forget the old rules. The difference lies in the apical tuft dendrites of layer 5b pyramidal neurons — not passive signal cables, but nonlinear "microcomputers" gated by NDNF inhibitory interneurons.

Citation: Maristany de las Casas et al. (2026). *Science* 392, eadx4358. DOI: 10.1126/science.adx4358 Institutions: Humboldt University of Berlin; Emory University Funding: Einstein Foundation, NIH, DFG SFB 1315

Key points

Relearning vs. execution dissociated: Optogenetic activation of NDNF interneurons (which suppress apical tuft activity) completely blocked mice's ability to *relearn* complex rules, while performance on already-learned simple tasks was unaffected. The tuft is a rule processor, not a motor executor.
New selection mechanism: Two-photon calcium imaging showed that NDNF activation suppressed calcium signals in the dendritic shaft while leaving spine signals intact — synapses still receive input, but dendritic integration is cut off.
Synapses cluster by rule: Glutamate imaging revealed that synapses on tuft dendrites are functionally clustered by rule context. Learning restructures the dendritic "functional map," storing different rules on physically separate branches — which explains why activating one rule does not erase another.
NDNF neurons as error monitors: They do not affect task execution; their activity correlates with errors during rule switches and rises only when relearning is needed — a gating system triggered by error signals rather than an external controller.

Background: from passive cable to active computer

Classical neuroscience treated neurons as point integrators. The Larkum group's work overturned this view over 17 years: calcium spikes in tuft dendrites (Science 2009), dendritic control of perception (Science 2016), layer 1 input controlling associative learning (Science 2020), and now dendritic computation as *necessary* for flexible learning (Science 2026). Pyramidal neurons have two input zones — basal dendrites encoding sensory features and apical tufts integrating contextual/feedback information — with the tufts performing nonlinear, parallel, branch-level computations gated by NDNF interneurons in a graded, not all-or-none, manner.

Implications for AI

Deep learning's catastrophic forgetting stems from global parameter updates; the brain stores rules in structurally separated dendritic branches with local plasticity.
A single pyramidal neuron may match a 2–5 layer neural network (Beniaguev et al., 2021), helping explain the cortex's efficiency.
Related engineering directions include Numenta's active dendrite networks, difference target propagation (Sacramento et al., 2018), and multi-compartment neuron models (Baronig & Legenstein, 2024).

Medical relevance

Reduced spine density in schizophrenia, abnormal dendritic pruning in autism, and spine abnormalities in Down syndrome all align with impaired dendritic computation and cognitive inflexibility. NDNF gating offers potential drug targets and suggests cognitive flexibility tests may index dendritic health.

References

Maristany de las Casas et al. (2026). Tuft dendrites in frontal motor cortex enable flexible learning. Science 392, eadx4358. DOI: 10.1126/science.adx4358
Larkum et al. (2009). Science 325, 756-760.
Takahashi et al. (2016). Science 354, 1587-1590.
Doron et al. (2020). Science 370, eaaz3136.
Beniaguev et al. (2021). Single Cortical Neurons as Deep Artificial Neural Networks. Neuron.
Iyer et al. (2022). Avoiding Catastrophe: Active Dendrites Enable Multi-Task Learning.
Baronig & Legenstein (2024). Context association in pyramidal neurons.
Naumann et al. (2024). Layer-specific control of inhibition by NDNF interneurons. bioRxiv.