The Neuron

An excitable cell that integrates thousands of inputs across a branching structure and emits a timed spike train.

The structure

A neuron has dendrites — a branching tree receiving contacts from thousands of other cells — a soma containing the nucleus, an axon that may extend a meter, and terminal boutons that contact the next cells. The membrane is studded with voltage-gated ion channels whose density varies systematically along the structure. Dendritic branches are not passive cables: they contain active conductances that let a single branch perform local nonlinear operations before anything reaches the soma. Cell types differ enormously in this geometry, and the taxonomy of cortical cell types is largely a taxonomy of shapes.

The function it dictates

The neuron integrates its inputs across space and time and emits an action potential when the membrane potential at the axon hillock crosses threshold. Because dendritic branches compute locally, a single neuron is better modeled as a small multilayer network than as a single summing unit. Information is carried in spike timing and rate, and the same average rate with different timing can mean different things downstream. Energy is a hard constraint throughout: the brain runs on roughly 20 watts, and spiking is expensive, which pushes the whole system toward sparse activity.

The mirror in AI

The artificial neuron — a weighted sum through a nonlinearity — abstracts this to its simplest possible form. The correspondence that holds: many weighted inputs, one integrated output, a threshold-like response, and learning stored in connection strengths rather than in units. ReLU sparsity is a loose echo of metabolic sparsity, though it arrived for gradient reasons rather than energetic ones. Attempts to close the gap — spiking neural networks, dendritic computation models — remain research directions rather than production practice.

The mirror in robotics

In robotics the closest match is not a controller but a sensor node in a distributed network: a unit that integrates local signals, applies a threshold, and transmits only when something has changed. Event-driven cameras, which report per-pixel brightness changes rather than frames, are a direct engineering translation of spike-based signaling, and they inherit its advantages — microsecond latency and very low power on static scenes.

Where the analogy breaks

Four differences matter. Artificial neurons emit a static scalar; biological ones emit timed spikes, and the timing carries information. Dendrites compute; the artificial model sums linearly. Neuromodulators such as dopamine and acetylcholine globally reconfigure how an entire circuit responds, with no equivalent in a standard network. And biological synapses update from locally available signals, whereas backpropagation requires a global error routed backward along the exact forward weights — a mechanism for which no convincing biological implementation has been found. The last of these is the deepest: our most successful learning algorithm is the one biology most clearly does not use.

Open questions

Is spike timing carrying information that rate codes discard, or is it an implementation detail?
How much of a cortical neuron's dendritic nonlinearity is computationally necessary?
What local learning rule approximates backpropagation well enough to explain cortical learning?

In Kiru Lab

The Artificial Neuron and Its Honest Limits Activations: Why the Nonlinearity Is Structural