The adaptive unit: learning without forgetting.
A machine that acts in the world keeps learning for as long as it runs — a new tool, a new floor, a new failure — and every new skill threatens the last. Train a standard network on a second task and it quietly overwrites the first; the field calls this catastrophic forgetting, and it is a large part of why today's models are trained once and then frozen. Biology does not freeze, and the difference is in the unit. A cortical neuron is not a single weighted sum: its dendrites each apply their own nonlinearity, so one neuron computes what an artificial network needs several layers to match — and that same structure lets different contexts recruit different dendrites. Put that inside an artificial unit and a network can route each task to its own sparse sub-network, learning the new without erasing the old. The Adaptive Unit studies the neuron as a small network, the continual learning its dendrites buy, and why the most capable version of the idea is also the one that fits the hardware we already have.
Run the same sequence of tasks through both networks and read the retention matrix: each cell is the accuracy on one task after the network has finished training up to a later one. The point-neuron network's early tasks redden and collapse toward chance as new tasks overwrite the shared weights — forgetting, in a picture. The dendritic-gated network's stay bright, because each task's context fires a different sparse set of dendrites, so their sub-networks barely overlap and barely interfere. Slide the number of active dendrites to watch sparsity trade capacity against interference. The second tab shows the reason the trick is available at all: toggle the NMDA nonlinearity and a single neuron's equivalent depth collapses from a small network to one layer.
You have probably seen the other kind of picture — a glowing tangle captioned "the brain of AI," a few dozen spheres standing in for a trillion, labels that mean nothing. This is the opposite of that. Below is a real network, drawn in full: every sphere is one neuron, every line one weight, and the light is the network actually computing the input you hand it. Drag it. Then switch on the dendrites and watch a task route to its own small set of neurons — the mechanism above, made visible.
Drag the scene to orbit it; drag the input pad to feed it a point and watch the activations flow. In dendritic mode, flip the task context and a different sparse sub-network lights up for the same input — one network, holding two skills in separate neurons. The forward pass runs on your device; nothing here is stock footage.
Why it matters: an agent that keeps its skills.
Continual, on-the-job learning is the capability embodied AI most lacks and most needs — a robot cannot be sent back to the factory to be retrained every time the world changes. The strongest evidence that dendritic gating delivers it was measured on a bench of simulated robot-manipulation tasks, exactly the setting this lab cares about.
In the field · the pieces are published and recent. A single cortical neuron's input-output takes a five-to-eight-layer network to match (Beniaguev, Segev & London, Neuron 2021) — and remove its NMDA dendritic nonlinearity and a single layer suffices, so the depth is the dendrite. Dendrite-structured networks match or beat conventional ones with far fewer parameters and less overfitting (Chavlis & Poirazi, Nature Communications 2025). And context-gated active dendrites hold 81% accuracy across 100 tasks learned in sequence, and about 88% across ten Meta-World robot tasks, where a standard network catastrophically forgets (Iyer et al., Frontiers in Neurorobotics 2022). Why hasn't the richer unit taken over? Because it is sparse and conditional, and the hardware of this era rewards dense matrix multiplies — the hardware lottery (Hooker, 2020). The frontier is quietly converging on it anyway: attention is input-dependent routing, mixture-of-experts is context-dependent sub-units, and gated (GLU) layers are multiplicative gating — the dendritic operation, re-expressed in a form a GPU runs fast. And the boldest test is now running: Pathway's Dragon Hatchling (BDH, 2025) makes the brain's mechanisms the architecture itself — a scale-free network of locally-interacting neuron particles whose working memory is Hebbian synaptic plasticity, whose activations are sparse, positive and demonstrably monosemantic, and which rivals GPT-2 at 10M-to-1B parameters. The sparse, local, interpretable-by-construction primitive, shown to scale.
See the mechanism — BDH holds its working memory in synapses it writes while it runs, by a Hebbian rule, and you can point at which synapse holds which fact. This is that idea at nano scale, on your device: teach it a fact and watch one synapse brighten, then ask a concept and it recalls its pair from a memory that was never trained in.
Take the course · PAI-135 →Open the full view ↗The neuron-as-deep-net code ↗
The paper
The argument set out in full: why the point neuron is an impoverished unit, what a dendritic segment buys once a context signal can select it, the measured continual-learning and robot-manipulation results with each attributed to the group that reported it, and the objection that matters most, which is that a primitive better suited to the problem can still lose to one better suited to the hardware. Section 6 lists what is unsettled, including the largest gap between this literature and a working robot: every result below supplies task identity, and an embodied machine is not told which task it is in.