Computational Neuroscience

Using mathematics and computer models to understand how the brain computes: from Hodgkin–Huxley and McCulloch–Pitts to Hopfield networks, reinforcement learning and large-scale simulations.

Brains process information, so the language of computation — mathematics, algorithms, simulations — is a natural tool for understanding them. Computational neuroscience builds models to explain how neural systems compute and to generate testable predictions.

Founding models

In 1943 the neurophysiologist Warren McCulloch and the logician Walter Pitts described idealised neurons as simple on/off units that sum weighted inputs and fire if a threshold is crossed, and showed networks of them could compute logical functions — a founding idea for both neuroscience theory and artificial intelligence. In 1952 the Hodgkin–Huxley model described the action potential with differential equations — a biophysically detailed model still in use. In the 1960s Wilfrid Rall showed how to model dendrites as electrical cables.

Levels of modelling

Models span scales. Biophysical models simulate ion channels and dendrites of single neurons in detail. Spiking network models simulate thousands to millions of simplified neurons ("integrate-and-fire"). Firing-rate and neural-field models describe average activity of populations. Normative models ask what the optimal solution to a problem would be — for example, Bayesian models of perception — and compare it with behaviour. David Marr's three levels — computational, algorithmic, implementational — help clarify what a model is meant to explain.

Landmark ideas

Hebbian learning formalises "fire together, wire together". In 1982 John Hopfield showed that a network with symmetric connections can store memories as stable states and retrieve them from partial cues (attractor networks) — work honoured with the 2024 Nobel Prize in Physics, shared with Geoffrey Hinton. Reinforcement learning algorithms, especially temporal-difference learning, turned out to match the reward prediction errors of dopamine neurons. Efficient coding theories explain sensory neurons' responses as compressing the statistics of natural scenes.

Large-scale simulation

Projects such as the Blue Brain Project (2005–2024) built detailed simulations of cortical tissue; whole-brain models combine connectome data with dynamics. Simulation alone, however, doesn't guarantee understanding — a perfect copy of a brain would be as hard to understand as the real thing.

Today

Computational neuroscience increasingly overlaps with machine learning: deep networks are used as models of the visual and language systems, and as tools for analysing huge datasets of neural recordings.