Dendrites as Microcomputers: Science Study Shows How the Brain Learns Flexibly While AI Forgets
> On May 7, 2026, Matthew E. Larkum's team at Humboldt University of Berlin published a study in *Science* revealing a stark contrast: mice can flexibly switch between complex rules, while AI models trained on the same tasks catastrophically forget the old ones. The difference lies in the apical tuft dendrites of layer 5b pyramidal neurons—passive-looking branches that are actually nonlinear "microcomputers," precisely gated by NDNF inhibitory interneurons. When optogenetics blocked this system, mice entirely lost the ability to relearn complex rules, yet execution of already-learned simple tasks was unaffected. This is direct evidence that active dendritic computation underlies flexible adaptation, and a key biological clue for solving catastrophic forgetting in AI.
Published: 2026-06-05 Source: Science 392, eadx4358 (2026) DOI: 10.1126/science.adx4358 Authors: Eduardo Maristany de las Casas, Kris Killmann, Moritz Drüke, Lukas Münster, Christian Ebner, Robert Sachdev, Dieter Jaeger, Matthew E. Larkum Institutions: Humboldt University of Berlin, Emory University Funding: Einstein Foundation, NIH, DFG SFB 1315
---
1. The Core Question: Why Can't AI Learn "Flexible Switching"?
Imagine driving into a country with completely reversed traffic rules. You would adapt within days without forgetting your old rules. A deep learning model, however, would overwrite old rules when learning new ones—catastrophic forgetting.
The brain doesn't need full retraining. The Larkum team's answer: computation happens not just in the whole neuron, but in its dendrites—especially the apical tuft dendrites, which act as independent "microcomputers."
---
2. Study Design: Separating "Execution" from "Relearning"
2.1 Behavioral experiments
Mice were trained on a rule-switching decision task: map cue signals (e.g., sounds) to left/right licking. The rules then reversed (sound A → right lick instead of left), forcing relearning rather than mere execution. This cleanly separates neural mechanisms of skilled execution from flexible learning.
2.2 The control switch: NDNF interneurons
NDNF (Neuron-Derived Neurotrophic Factor) interneurons are inhibitory cells in cortical layer 1, sitting right beside pyramidal apical tufts. Using optogenetics, the team activated NDNF neurons to suppress apical tuft activity—effectively pausing the microcomputers.
---
3. Four Core Findings
Finding 1: Blocking dendritic computation abolishes relearning but not execution
With NDNF neurons activated:
- Learned simple tasks: performance unaffected
- Relearning complex rules: completely blocked
- Suppressed calcium signals in the dendritic shaft
- Left spine calcium signals unaffected
- Do not affect task execution
- Their activity correlates highly with errors during rule switches
- Become active only when relearning is needed
- 2009: apical tufts generate independent calcium spikes (Science)
- 2016: dendritic calcium spikes modulate perception (Takahashi et al., Science)
- 2020: layer 1 input (including NDNF neurons) controls associative learning (Doron et al., Science)
- 2026: dendritic computation is necessary for flexible learning
- NDNF quiet → tufts compute → context integrated → flexible learning possible
- NDNF active → tufts suppressed → context blocked → only existing rules executable
- Catastrophic forgetting stems from global parameter updates; brains store rules in locally separated dendritic branches.
- A pyramidal neuron runs thousands of parallel branch computations—natural parallelism that AI approximates only artificially (EWC, MoE).
- Modeling work (Beniaguev et al., 2021) shows a single cortical neuron requires a ≥2-hidden-layer MLP to reproduce—helping explain the brain's efficiency.
- Relevant approaches: Numenta's active dendrite networks, Difference Target Propagation (Sacramento et al., 2018), multi-compartment neuron models (Baronig & Legenstein, 2024).
- Key insight: gating is triggered by error signals themselves, not an external controller.
- Schizophrenia: reduced dendritic spine density, apical dendrite atrophy
- Autism: excessive branch growth or abnormal pruning
- Down syndrome: abnormal spines linked to motor and intellectual impairment
- Apical tufts are nonlinear "microcomputers," not passive wires
- NDNF interneurons gate which contextual information enters dendritic computation
- Blocking tufts preserves execution but abolishes relearning
- Synapses cluster by rule on dendritic branches—learning reshapes the branch functional map
- 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). Synaptic Integration in Tuft Dendrites of Layer 5 Pyramidal Neurons. Science 325, 756-760.
- Takahashi et al. (2016). Active cortical dendrites modulate perception. Science 354, 1587-1590.
- Doron et al. (2020). Perirhinal input to neocortical layer 1 controls learning. 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 in Dynamic Environments.
- Baronig & Legenstein (2024). Context association in pyramidal neurons through local synaptic plasticity in apical dendrites.
- Naumann et al. (2024). Layer-specific control of inhibition by NDNF interneurons. bioRxiv.
Apical tufts are not motor executors—they are rule processors for adjusting behavior in new contexts.
Finding 2: Calcium signals suppressed in the dendritic shaft but not spines—a new information selection mechanism
Two-photon calcium imaging showed that NDNF activation:
Individual synaptic inputs still arrive, but dendritic integration is cut off. This suggests a new selection mechanism: information can pass at the synapse level while being excluded from dendritic-level computation.
Finding 3: Synapses cluster on dendrites by rule context—learning changes structure, not just weights
Glutamate imaging revealed that synapses on apical tufts are functionally clustered by rule context: synapses for rule A occupy some branches, rule B synapses occupy others.
Learning therefore doesn't just adjust synaptic weights—it reshapes the dendritic branch "functional map." Different rules are assigned to physically separated dendritic compute units, so activating one does not erase the other.
Finding 4: NDNF neurons are "error monitors"
NDNF interneurons:
They are monitors of cognitive flexibility—detecting errors and triggering dendritic recomputation.
---
4. Biological Basis of Dendritic Computation
4.1 Collapse of the traditional view
Classical neuroscience treated neurons as point integrators with "passive wire" dendrites. The Larkum lab overturned this across a 17-year arc:
4.2 Dual computation in pyramidal neurons
| Region | Location | Inputs | Function | |--------|----------|--------|----------| | Basal dendrites | Near soma | Lower cortical/sensory | Encode external features | | Apical tuft | Layer 1, far from soma | Higher cortical, context/feedback | Nonlinear contextual integration |
The tuft generates calcium-mediated local events that can modulate somatic output, trigger local plasticity, and support parallel independent computation across branches.
4.3 NDNF gating
The gate is graded, not all-or-none.
---
5. Implications for AI
---
6. Medical Significance
If dendritic computation underlies flexible learning, dendritic abnormalities should impair cognitive flexibility:
NDNF gating offers potential drug targets (GABA receptor modulators, compounds enhancing dendritic calcium signals), and cognitive flexibility tests may serve as a proxy for dendritic computation health.
---
7. Open Questions
1. How does dendritic computation avoid interfering with old knowledge (molecular mechanism)? 2. Can NDNF gating be adapted into trainable AI gating mechanisms? 3. What does the extreme energy efficiency of dendritic computation imply for neuromorphic computing? 4. Does human (prefrontal) dendritic computation differ qualitatively from mouse? 5. Can dendritic-style local learning be integrated into deep learning without sacrificing training efficiency? 6. Is dendritic computation also the biological basis of creativity?
---
8. Summary
Using optogenetics, calcium imaging, and glutamate imaging, the Larkum team showed that active dendritic computation is the biological foundation of flexible adaptation:
For AI, the root of catastrophic forgetting may be the absence of local dendritic computation and structural separation; brain-inspired systems need "point neurons" upgraded to multi-compartment neurons with local learning and dynamic gating.
The open question: while AI researchers train billion-parameter models with global backpropagation, biological brains run thousands of parallel computations inside a single neuron. That gap may be the most essential distance between current AI and genuine intelligence.
---