The Wilson-Cowan Model

Computational Neuroscience

HH gives you one neuron in obsessive detail. LIF lets you simulate millions of them. For most questions about cortical dynamics, even LIF is too much resolution — you don't want to track individual neurons, you want their average firing rates. Wilson-Cowan is the canonical population-rate model: two coupled ODEs for the mean activity of excitatory and inhibitory neurons. Published by Hugh Wilson and Jack Cowan in 1972, still the workhorse for thinking about population-level cortical dynamics.

The conceptual move is the same one statistical mechanics makes: stop tracking individuals, track averages. For cortex this works because neighboring neurons of the same type mostly fire similar amounts at any given moment — the population rate is a meaningful summary, even though individual spike times are not.

The setup

Divide cortical neurons into two populations. The excitatory population $E$ — mostly pyramidal cells, glutamatergic, about 80% of cortex. The inhibitory population $I$ — mostly interneurons, GABAergic, about 20%. At any moment each population has an average firing rate normalized to $[0, 1]$ (the fraction of its neurons currently active). The two variables of interest are $E (t)$ and $I (t)$ .

Each population receives input from itself (recurrent connectivity), from the other population (E recruits I, I suppresses E), and from outside (sensory input, top-down attention, anything not in the model). The instantaneous rate of each population relaxes toward a target rate that depends on its total synaptic drive.

The equations

The Wilson-Cowan equations are:

τ_{E} \frac{d E}{d t} τ_{I} \frac{d I}{d t} = - E + F (w_{EE} E - w_{E I} I + P_{E}) = - I + F (w_{I E} E - w_{II} I + P_{I})

Each equation is a leaky integrator (the $- E$ term pulls $E$ back toward zero with time constant $τ_{E}$ ) plus a forcing term (the $F (\cdot)$ drives $E$ toward a target activation given current total drive). Same shape as LIF's subthreshold equation, but at the population level instead of the single-neuron level.

$F$ is a sigmoidal activation function — usually a logistic:

F (x) = \frac{1}{1 + exp ( - ( x - θ ) / k )}

$F$ bounds the population rate in $[0, 1]$ — no negative rates, no rates greater than 1. The threshold $θ$ sets where activation kicks in; the slope $k$ sets how sharply. Below $θ$ , $F$ is near zero; above $θ$ , it saturates near one. The sigmoid is what makes Wilson-Cowan nonlinear, and the nonlinearity is the whole thing — it's what produces oscillations, bistability, and excitability.

The four $w$ parameters are coupling strengths, all positive: $w_{EE}$ (E excites itself), $w_{E I}$ (I suppresses E), $w_{I E}$ (E excites I), $w_{II}$ (I suppresses itself). The sign structure is fixed by biology — E neurons release glutamate, I neurons release GABA. You can't have an excitatory neuron that suppresses its target in this framework.

$P_{E}$ and $P_{I}$ are external inputs to each population. Time constants $τ_{E}$ and $τ_{I}$ are typically a few milliseconds, with $τ_{I} > τ_{E}$ — inhibition is slower than excitation, mostly because GABA-A receptors decay more slowly than AMPA receptors. This time-constant asymmetry turns out to be essential for the oscillating regime.

The four regimes

Wilson-Cowan exhibits four qualitatively distinct regimes, separated by bifurcations in parameter space. The simulator below walks through each one.

P_E0.00

Weak external drive. Both populations settle to a low-activity fixed point. No oscillations, no bistability — just a leaky filter waiting for input.

Resting. Both rates settle to a low fixed-point value. The population is a low-pass filter waiting for input. Important but unexciting on its own — it's the regime that holds when external drive is below threshold.

Oscillations. Strong $E \leftrightarrow I$ coupling with $τ_{I} > τ_{E}$ produces a Hopf bifurcation: the rest fixed point loses stability and the trajectory spirals out onto a limit cycle. $E$ ramps up, recruits $I$ , $I$ suppresses $E$ , both rates fall, $I$ decays back faster than $E$ can recover the asymmetry, $E$ climbs again. The frequency depends on the time constants — typically 20–80 Hz for cortical parameters. This is the canonical mechanism behind the cortical gamma rhythm.

Bistability. Strong recurrent $E \to E$ excitation creates two stable fixed points: a low-activity rest, and a high-activity "persistent" state. A saddle point sits between them. Small perturbations decide which basin the trajectory ends up in. The high state is one canonical mechanism for working memory — a brief stimulus pushes the system into the high state, which is then maintained until reset by other input. Click the "kick E" button in the simulator under this preset to try crossing the separatrix.

Excitable. Single stable fixed point, but a perturbation triggers a big detour through phase space before relaxing back. Same machinery as a neuron's action potential, lifted to the population level — a brief "population spike" in response to a kick. The kick has to be big enough to push past the unstable manifold; smaller kicks just decay back without exciting the transient.

Saturation. Very strong external drive holds both populations near their maximum rate. The sigmoid's upper plateau dominates the dynamics. No interesting transients — just a clamped high state. Wilson-Cowan deliberately excludes this regime in most applications by keeping inputs in a reasonable range.

The phase plane

The phase plane in the simulator is where the geometry lives. Two key curves: the E-nullcline (where $d E / d t = 0$ ) and the I-nullcline (where $d I / d t = 0$ ). Their intersections are the fixed points of the system.

Setting $d E / d t = 0$ in the first equation and solving for $I$ in terms of $E$ :

I = \frac{w _{EE} E - F ^{- 1} ( E ) + P _{E}}{w _{E I}}

And from $d I / d t = 0$ in the second equation, solving for $E$ in terms of $I$ :

E = \frac{w _{II} I + F ^{- 1} ( I ) - P _{I}}{w _{I E}}

Where $F^{- 1} (y) = θ + k ln (y / (1 - y))$ is the logit function. Both nullclines come out S-shaped, and the number and stability of their intersections classifies the regime. One stable intersection: resting state. An unstable spiral with a surrounding limit cycle: oscillations. Three intersections (two stable + one saddle): bistability. The transitions between these regimes happen at bifurcations — Hopf (rest → oscillations), saddle-node (one fixed point → three).

The phase-plane structure here is the same machinery that classifies any 2D dynamical system. Wilson-Cowan is the textbook example because the biology is clean (two populations, sign-fixed couplings), the regimes are pedagogically clear, and the bifurcations are accessible by hand if you want to compute them.

What Wilson-Cowan captures

Read each regime as a cortical phenomenon.

Cortical rhythms. The oscillating regime gives you gamma-band oscillations (~40 Hz) from E-I interaction with appropriate time constants. Variants with slower variables added produce alpha (10 Hz) and theta (4–8 Hz). Almost every dynamical account of cortical rhythms uses Wilson-Cowan or a close cousin — Jansen-Rit, neural mass models, mean-field reductions of spiking networks all share its skeleton.

Working memory and persistent activity. The bistable regime is the canonical mechanism for sustained activity that outlasts its stimulus. Real cortical "memory cells" do this: they fire continuously throughout a delay period, then return to baseline after the response. Wilson-Cowan gives the dynamical-systems explanation — the network has been pushed into the high-activity basin and stays there.

Decision-making. Two coupled Wilson-Cowan systems with mutual inhibition produce winner-take-all dynamics. This is the basis of standard models of perceptual decision-making (Wong & Wang 2006 is the canonical reference).

Avalanches and critical dynamics. The excitable regime, with appropriate noise, produces avalanche-like activity patterns matching what's measured in real cortical recordings. Connects to theories of self-organized criticality in cortex.

What Wilson-Cowan abstracts away

Wilson-Cowan trades resolution for tractability. Individual neurons disappear into the population average — there's no spike timing, no individual neuron voltage trajectory, no heterogeneity within the population. Synapses are absorbed into the $w$ couplings, so synaptic plasticity has to be re-added by making those couplings dynamic. Spatial structure (cortical columns, retinotopy, layer-specific connectivity) is gone — Wilson-Cowan is one node, not a network of nodes.

For network-of-nodes models, Wilson-Cowan is the per-node dynamics, with additional coupling between nodes representing inter-region connections. This is the structure of modern whole-brain models — the Virtual Brain project uses Wilson-Cowan or close relatives at each parcellated brain region, coupled by an anatomical connectivity matrix. So even when the "neurons" are gone, the equations are still there.

Closing

Wilson-Cowan is the canonical entry to dynamical-systems neuroscience. The phase-plane picture you build here — nullclines, fixed points, bifurcations between regimes — is the same picture that classifies FitzHugh-Nagumo (the 2D reduction of HH, a natural next post), Wong-Wang decision models, mean-field reductions of Hopfield networks, and whole-brain coupled-oscillator systems. The biology decides the sign structure of the couplings and gives you reasonable parameter ranges; the math underneath is general 2D dynamical systems, and the toolkit you build here transfers everywhere.