What a Single Neuron Can Compute

For a long time, the standard picture of a neuron was a tidy one. Inputs arrive, they are added up, and if the total crosses a threshold the neuron fires. This picture, a weighted sum followed by a switch, is the ancestor of the artificial neurons used in machine learning today.

It is a useful simplification. It is also, increasingly clearly, an incomplete one.

The branches matter

A real neuron receives most of its inputs on its dendrites, branching structures that can span hundreds of micrometres. Those branches are not passive wires. Many contain ion channels that let them generate local electrical events of their own.¹

Because of this, inputs that arrive close together on the same branch can combine in ways that are more than additive. A branch can behave like a small processing unit, responding strongly only when several inputs coincide.

A single neuron may be less like a switch and more like a small network folded into one cell.

Why it changes the picture

If individual branches can compute, then the computational power of a neuron depends on how its inputs are arranged, not only on how many there are. Two neurons receiving the same inputs could respond quite differently depending on where those inputs land.

This has consequences for how we model the brain. Simple point-neuron models remain valuable, but researchers now also build models in which each neuron contains several interacting compartments.

An open frontier

How much of the brain's capability depends on this fine-grained, within-cell computation is still being worked out. It is a reminder that even the basic units of the nervous system hold more structure than early diagrams suggested, and that the right level of simplification depends on the question being asked.