Charlot Lab · The Adaptive Unit

Learning without forgetting

An embodied agent has to learn a new skill without unlearning the last one. A standard network can't: train it on a second task and it overwrites the first — catastrophic forgetting. A dendritic unit routes each task to its own sparse sub-network, so the tasks stop colliding. Watch both learn the same sequence, live, on your device.

Retention matrix · accuracy on each task, after training each task in sequence

Point-neuron network

One shared network. Every task rewrites the same weights.

mean over all tasks:

Dendritic-gated network

Each task's context activates a sparse set of dendrites → its own sub-network.

mean over all tasks:

Each cell = test accuracy on task row after the network has finished training up to task column. The point network's early rows go dark as later tasks overwrite them. The dendritic network's stay bright: different tasks fire different dendrites, so their sub-networks barely overlap and barely interfere. Diagonal = fresh learning; below the diagonal = memory.

Decision boundary · dendritic net, task 1
active dendrites (k) 6 of 24
A layer-5 pyramidal cell
Drive its dendritic branches
equivalent feed-forward net to match this cell's input→output:
7 layers deep

A point neuron is one weighted sum and a threshold — the black dot at the soma. But a real neuron's branches each apply their own nonlinearity first (NMDA channels), so the cell computes a layered function. Toggle NMDA off and the branches go linear: the whole cell collapses to what a single layer can do. That switch is the entire difference between "a neuron is a dot" and "a neuron is a small network."