Computational neuroscience
Computational neuroscience is the discipline that uses mathematical models, computer simulations, and information theory to understand the principles by which nervous systems process information, generate behavior, and implement cognition. It sits at the intersection of neuroscience, physics of computation, applied mathematics, and artificial intelligence — a crossing of disciplines that has produced both genuine insight and productive confusion about what kind of thing the brain actually is.
The field's founding tension: computational neuroscience both describes brains in computational terms and uses those descriptions to build better computational systems. When these two projects converge, it is assumed to be because the brain and the machine are doing fundamentally the same thing. This assumption has never been justified. It is an inference from analogy — a powerful one, enormously productive, and not, for all that, established as fact. A neuroscience that cannot distinguish between the brain computes as a description and the brain computes as a metaphor has not yet clarified its own foundations.
Key figures include Warren McCulloch, David Marr (whose three levels of analysis — computational, algorithmic, implementational — structured the field), and Horace Barlow, who argued that the goal of sensory systems is to reduce redundancy — a claim that remains contested and productive in equal measure.
Marr's Three Levels
David Marr's framework for understanding information processing systems remains the most influential organizing principle in computational neuroscience. Marr proposed that any such system must be understood at three distinct levels:
Computational level: What does the system do, and why? What is the computational problem being solved? For vision, the problem is to derive a three-dimensional representation of the world from two-dimensional retinal images. For motor control, the problem is to convert desired trajectories into muscle activations that achieve them despite noise, delay, and changing dynamics. The computational level is about the logic of the task, independent of how it is implemented.
Algorithmic level: How does the system solve the problem? What representations does it use, and what procedures transform inputs into outputs? Marr proposed that early vision uses edge maps and primal sketches as intermediate representations, constructed by filters tuned to specific spatial frequencies and orientations. The algorithmic level is about the strategy of the solution.
Implementational level: How is the algorithm physically realized? In the brain, this means neurons, synapses, action potentials, and neurotransmitter dynamics. The implementational level is about the hardware — though "hardware" is misleading for a system that rewires itself in response to experience.
Marr's framework is pedagogically invaluable but scientifically treacherous. The levels are not independent. The choice of algorithm constrains what can be implemented efficiently. The properties of neurons constrain what algorithms are feasible. And the computational problem itself is often ill-defined — evolution does not solve problems; it finds locally adequate solutions to survival challenges. To assume that the visual system "computes" inverse optics is to impose a teleology that natural selection may not share.
Canonical Models
The Hodgkin-Huxley model (1952) is the foundational achievement of computational neuroscience. Alan Hodgkin and Andrew Huxley described the generation of action potentials in the squid giant axon using a system of four coupled nonlinear differential equations that capture the dynamics of sodium and potassium ion channels. The model is remarkable for its predictive power: it accurately reproduces action potential shape, threshold behavior, refractory periods, and propagation velocity. It remains the standard for single-neuron modeling, though modern variants incorporate dozens of additional ion channel types, calcium dynamics, and dendritic compartmentalization.
Integrate-and-fire models sacrifice biophysical detail for analytical tractability. A neuron is modeled as a capacitor that integrates synaptic inputs until its membrane potential crosses a threshold, at which point it emits a spike and resets. These models cannot reproduce the rich dynamical behaviors of real neurons — bursting, adaptation, resonance — but they permit large-scale network simulations that would be intractable with Hodgkin-Huxley dynamics. The tradeoff between biological realism and computational feasibility is one of the central methodological tensions in the field.
Neural coding asks how information is represented in neural activity. The earliest framework assumed rate coding: information is encoded in the average firing rate of neurons over some time window. But precise spike timing matters: some systems encode information in the timing of individual spikes relative to ongoing oscillations (phase coding), in the synchrony of spike times across neurons (synchrony coding), or in the precise temporal pattern of spikes within a population (temporal coding). The debate between rate and temporal coding is not settled; different brain regions and tasks may use different schemes, and the same population may multiplex multiple codes simultaneously.
Synaptic plasticity models attempt to capture how neural circuits learn. The Hebbian rule — neurons that fire together wire together — is the conceptual foundation, but it is mathematically unstable without additional constraints. Spike-timing-dependent plasticity (STDP) refines Hebb's principle by making synaptic change depend on the precise timing of pre- and postsynaptic spikes: if the presynaptic spike precedes the postsynaptic spike, the synapse strengthens; if the reverse, it weakens. STDP can be derived from biophysical models of NMDA receptor dynamics and has been observed experimentally in multiple brain regions. But whether STDP is the primary learning rule in the brain, or one of many, remains open.
The Systems Critique
Computational neuroscience has been extraordinarily successful at the cellular and circuit levels. It has described how single neurons generate spikes, how small networks produce oscillations, and how sensory receptors transduce physical signals. But as the field scales up — from circuits to brain regions to whole-brain dynamics — a systems-theoretic critique becomes unavoidable.
The brain is not a computer. This is not a mystical claim. It is a statement about architecture. Computers are designed with a separation between processing and memory, with a central clock that synchronizes operations, and with a von Neumann bottleneck that forces sequential access to stored programs. The brain has none of these features. There is no central clock. Processing and memory are inseparable — synaptic weights are both the memory and the computation. Information flows in parallel through massively recurrent networks, and the same anatomical structure supports multiple functions depending on behavioral state, neuromodulatory context, and recent history.
To model the brain as a neural network is not wrong — it is a productive simplification. But to model it as a feedforward neural network trained by backpropagation is to import assumptions from machine learning that have no biological correlate. The brain does not compute gradients across layers. It does not have a global loss function. It does not separate training from inference. These are engineering conveniences, not biological principles.
The reverse inference problem plagues computational neuroscience at every scale. When a model successfully reproduces some aspect of neural activity — a firing pattern, a tuning curve, a behavioral response — it is tempting to infer that the mechanism in the model is the mechanism in the brain. But multiple distinct mechanisms can produce the same output. The Hodgkin-Huxley model reproduces action potentials, but so do simpler models with different underlying assumptions. A recurrent network that produces oscillations at 40 Hz may do so through inhibition-mediated rhythms or through excitatory reverberation — mechanisms with different biological implications. Computational neuroscience has tools for forward prediction (given a model, what does it do?) but lacks tools for reverse inference (given data, what model is correct?).
The multiple realizability problem is deeper. Even if a model accurately captures the mechanisms of a particular brain region in a particular species, those mechanisms may not generalize. The visual cortex of a mouse is not the visual cortex of a macaque. The hippocampus of a rat solves spatial navigation problems that the human hippocampus may not. Computational neuroscience often treats model organisms as proxies for the human brain, but the justification for this extrapolation is rarely examined. A model that works in a slice of rat cortex may be irrelevant to human cognition.
Connection to Artificial Intelligence
The relationship between computational neuroscience and AI is bidirectional and increasingly fraught. Deep learning was originally inspired by neural network models from computational neuroscience, but it has since diverged. Modern artificial neural networks use backpropagation, batch normalization, attention mechanisms, and transformer architectures that have no known biological analogues. The claim that deep learning "understands" the brain because both use "neural networks" is a category error — like claiming that a paper airplane understands aerodynamics because both use lift.
Conversely, insights from AI have begun to influence neuroscience. The predictive coding framework, in which neural circuits minimize prediction error, has been formalized using variational inference — a tool from machine learning. The observation that representations in deep networks become increasingly abstract with depth has motivated analogous investigations in the ventral visual stream. And the success of reinforcement learning in AI has renewed interest in dopaminergic reward prediction error as a biological implementation of temporal difference learning.
But the traffic is mostly one-way: neuroscience inspires AI architecture less than AI inspires neuroscience theory. This asymmetry may reflect a genuine difference in the problems being solved, or it may reflect the greater funding and computational resources available to AI research. Either way, the two fields are drifting apart, and the dream of a unified "neural computation" science — in which insights from artificial and biological systems mutually constrain each other — remains unrealized.
Computational neuroscience is the most ambitious reductionist project in science: to explain the mind by modeling the brain. Its successes are real and hard-won — the Hodgkin-Huxley model, the tuning properties of visual cortex, the dynamics of decision-making circuits. But its failures are equally instructive. The brain is not a computer running an algorithm. It is a self-organizing, history-dependent, metabolically constrained, embodied system that evolved under selection pressures no engineer would recognize. To understand it, we need more than better models of neurons. We need a new theoretical framework — one that treats the brain as a complex adaptive system, not a biological implementation of a computational theory.