Computational Neuroscience

Computational neuroscience is the branch of neuroscience that treats the brain as an information-processing system and tries to explain how it works using mathematics, models, and simulation.

10 frameworks, from 1943 to today · Neuroscience · Natural Sciences

still a working view ended · each leads to its entry below

Overview

Computational neuroscience is the branch of neuroscience that treats the brain as an information-processing system and tries to explain how it works using mathematics, models, and simulation1. Its practitioners ask questions like: How does a neuron turn a stream of tiny electrical inputs into an output spike? How does a network of neurons store a memory or represent a face? What computation is the visual cortex actually performing? What is at stake is whether the brain's most remarkable abilities—perception, memory, learning, action—can be understood as the operation of a physical system, and whether building models of that system can tell us something that experiments alone cannot.

The field is unusual in that it has always been a meeting point. Physiologists, physicists, engineers, mathematicians, and computer scientists have all contributed, and the history is one of ideas migrating between the brain and the machine in both directions.

The neuron as a circuit element: from the 1940s

The founding move was to treat the neuron as a computational unit. In 1943, Warren McCulloch and Walter Pitts published a paper showing that idealized neurons—binary units that fire when their summed inputs cross a threshold—could, in principle, compute any logical function. This was a mathematical abstraction, not a claim about real anatomy, but it established the idea that nervous tissue could be described as a logic circuit.

The next step came from physiology. In the early 1950s, Alan Hodgkin and Andrew Huxley worked out the ionic mechanism of the action potential in the squid giant axon, measuring how sodium and potassium currents change with voltage and membrane permeability. Their model, published in 1952, is a set of differential equations describing how the membrane potential evolves2. It remains the benchmark for biophysical realism: it predicts the shape and timing of spikes from the underlying ion channels. The Hodgkin–Huxley equations showed that a neuron's electrical behavior could be derived from measurable physical properties, and they set the standard for what a "mechanistic" model should look like.

At the same time, the question of what neurons represent was being posed. In the late 1950s and early 1960s, David Hubel and Torsten Wiesel recorded from the visual cortex of cats and found cells that responded to edges at particular orientations, and others that responded to movement in a preferred direction. These were the first "feature detectors"—neurons whose firing seemed to encode a specific property of the stimulus. The finding raised a computational question: how could such selectivity arise from the wiring of the visual pathway? In 1962, Hubel and Wiesel proposed a model in which simple cells combine inputs from center-surround retinal ganglion cells, and complex cells combine inputs from simple cells. This was one of the first attempts to explain a cortical response property as the result of a feedforward network.

Learning rules and the first networks

If neurons are computing units, how do they get wired up? In 1949, Donald Hebb proposed that when one neuron repeatedly helps fire another, the connection between them strengthens34. This "Hebbian" rule—cells that fire together wire together—gave a plausible mechanism for learning and memory, and it became a cornerstone of the field34. It was a local rule: it used only information available at the synapse.

In 1958, Frank Rosenblatt built the perceptron, a trainable pattern-recognition machine inspired by the neuron. The perceptron learned by adjusting its weights according to errors, and it could classify simple patterns. The idea that a network could learn from examples was powerful, and it attracted both excitement and criticism. In 1969, Marvin Minsky and Seymour Papert published a book showing that a single-layer perceptron could not solve certain simple problems (like exclusive-or), which dampened enthusiasm for neural networks for more than a decade56. The limitation was real: the perceptron had no hidden layer, and no one yet had a good way to train one.

The solution came in the 1980s6. In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams popularized the backpropagation algorithm, which computes how much each weight in a multi-layer network contributed to the output error and adjusts it accordingly. Backpropagation made it possible to train networks with hidden layers, and it became the workhorse of connectionist modeling. The algorithm is not biologically plausible in detail—it requires a backward pass of error signals that real neurons do not obviously have—but it demonstrated that layered networks could learn complex mappings, and it gave computational neuroscientists a tool for asking what kinds of representations a network could learn from data.

The brain as an associative memory

A parallel tradition asked how networks of neurons could store and retrieve memories. In 1982, John Hopfield published a model of a network of binary neurons with symmetric connections, showing that such a network could act as a content-addressable memory: if you present a partial or noisy version of a stored pattern, the network settles into the full pattern7. These "attractor networks" (the stored patterns are attractors of the network's dynamics) gave a concrete, analyzable model of associative memory87. The Hopfield model was a physicist's approach—it borrowed tools from statistical mechanics—and it inspired a generation of work on how neural circuits might implement memory, including later models of grid cells and head-direction cells as attractor networks.

Vision as computation: Marr's levels

In the late 1970s, David Marr, working at MIT, proposed a framework for understanding vision that reshaped how computational neuroscientists thought about their subject. Marr argued that any information-processing system should be understood at three levels: the computational level (what problem is being solved and why), the algorithmic level (what representations and procedures are used), and the implementation level (how the hardware carries it out). His 1982 book Vision applied this framework to early visual processing, proposing specific algorithms for edge detection and stereopsis. Marr's levels gave the field a vocabulary and a justification for building models that were not necessarily biophysically detailed: a model could be valuable if it captured the right computation, even if its implementation was abstract. The framework also made clear that "what the brain computes" was a legitimate scientific question, distinct from "how the neurons are wired."

Neural coding: what do spikes mean?

As recording techniques improved, a central question became how information is represented in the activity of populations of neurons. In the 1970s and 1980s, work on the visual system and on the motor cortex suggested that information could be read out from the firing rates of many neurons. Apostolos Georgopoulos and colleagues, recording from motor cortex in the 1980s, found that the direction of arm movement could be predicted from the population activity of many neurons, each of which was broadly tuned. This "population vector" idea suggested that the brain might represent movement direction as a distributed pattern across a population, rather than in the firing of any single cell9.

The 1990s and 2000s brought more sophisticated statistical tools. Researchers began to ask how much information a spike train carries, using information theory, and how correlations between neurons affect what a population can represent. The debate between "rate coding" (information in the number of spikes per unit time) and "temporal coding" (information in the precise timing of spikes) remains open, and the answer likely depends on the brain region and the task.

The Bayesian brain and predictive processing

A major development of the 1990s and 2000s was the idea that the brain performs probabilistic inference. The Bayesian framework treats perception as the problem of inferring the causes of sensory input, given prior expectations and noisy evidence10. In the 1990s, researchers like David Knill and Alexandre Pouget showed that human behavior in sensory tasks often matches the predictions of Bayesian optimality. The idea was extended by Karl Friston and others into "predictive processing": the brain is constantly generating predictions about its input and using the mismatch (prediction error) to update its internal model. This framework has been influential because it offers a unified account of perception, learning, and action, and because it connects to the error-driven learning ideas of the connectionists. It remains a subject of active debate, particularly about how literally the brain implements Bayesian computations.

Deep learning and the brain

The most recent turning point is the rise of deep learning. Since the early 2010s, deep neural networks—networks with many layers, trained on large datasets using backpropagation—have achieved dramatic successes in tasks like image recognition and language processing. These networks were not designed as models of the brain, but computational neuroscientists quickly noticed that their internal representations often resemble those found in the brain. In 2014, Daniel Yamins and colleagues showed that deep networks trained on object recognition produced representations in their higher layers that predicted neural responses in the ventral visual stream better than earlier models. This has led to a productive, if contested, research program: using deep networks as "model organisms" for brain computation, and asking which architectural features are necessary to reproduce neural response properties1. The debate about whether backpropagation is biologically plausible, and what alternative learning rules might work, is one of the liveliest in the field today.

The present

Computational neuroscience today is a broad church. Biophysical modelers build detailed compartmental models of single neurons and small circuits, using the Hodgkin–Huxley framework and increasingly detailed anatomical data. Network modelers study how populations of neurons produce dynamics like oscillations, attractor states, and sequential activity. Statistical neuroscientists develop methods for decoding neural activity and for testing models against data. And a growing community uses deep learning both as a tool for analyzing neural data and as a source of hypotheses about brain computation1.

The field's central questions remain open. How does the brain learn without a global error signal? What is the right level of abstraction for a model of a given brain region? How do we connect the dynamics of a network to the computations it performs? And perhaps most fundamentally: what would it mean to have a model that truly explains a piece of the brain, rather than merely predicting its activity? These questions are the direct descendants of the ones McCulloch and Pitts, Hodgkin and Huxley, and Marr were asking, and they are still being answered.

References

  1. Computational neuroscience (introduction, Cognition, discrimination, and learning), Wikipedia
  2. Hodgkin–Huxley model, Wikipedia
  3. Hebbian theory, Wikipedia
  4. Spike-timing-dependent plasticity (History), Wikipedia
  5. Connectionism, Wikipedia
  6. Neural network (History), Wikipedia
  7. Hopfield network (introduction, History), Wikipedia
  8. Attractor network, Wikipedia
  9. Neural coding (Population coding), Wikipedia
  10. Bayesian approaches to brain function, Wikipedia

The frameworks of Computational Neuroscience

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1943-Present
  2. 1949-Present

    Hebbian theory

    How do connections between neurons change with experience?

    Fly to its tower in Nousfarer

  3. 1952-Present

    Biophysical compartmental modeling

    How do ion channels produce the action potential?

    Fly to its tower in Nousfarer

  4. 1962-Present
  5. 1970-Present

    Neural coding

    What information do spike trains carry?

    Fly to its tower in Nousfarer

  6. 1977-Present

    Marr's levels of analysis

    At what levels should we explain an information-processing system?

    Fly to its tower in Nousfarer

  7. 1982-Present

    Attractor network

    How can networks store and recall memories?

    Fly to its tower in Nousfarer

  8. 1986-Present
  9. 1990-Present

    Bayesian brain and predictive processing

    Does the brain infer causes probabilistically?

    Fly to its tower in Nousfarer

  10. 2014-Present

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Artificial neural network modelingHebbian theoryBiophysical compartmental modelingFeature-detector and hierarchical fee…Neural codingMarr's levels of analysisAttractor networkError-driven learning and backpropaga…Bayesian brain and predictive process…Deep learning as a model of brain com…

Beyond the field

Where Computational Neuroscience's ideas came from in other fields, and what they led to there: the strongest 12 of 66 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Computational Neuroscience is a planet in Nousfarer

Its city has a tower for each of these 10 frameworks. Land, walk in, and climb one concept at a time.

Land on Computational Neuroscience

Related fields

Fields its ideas are linked to

More in Neuroscience

See where Computational Neuroscience sits on the Map of Human Knowledge