CALLIOPE · BCOM

This work presents a unified theoretical framework—termed the "Rosetta Stone"—that organizes the major families of neural mass models (NMMs) along a six-rung ladder of increasing biological complexity, revealing that all such models share a common push–pull dynamical motif between excitatory and inhibitory variables. The framework spans from the undamped harmonic oscillator and Kuramoto phase model through the Stuart–Landau oscillator, Wilson–Cowan equations, and second-order synaptic NMMs (Jansen–Rit family), up to next-generation mean-field models derived exactly from quadratic integrate-and-fire neurons. At each rung, explicit mathematical transformations expose which assumptions are introduced or relaxed, enabling principled translation of results across formalisms rather than ad hoc model selection. The treatment further standardizes how external inputs—including transcranial electrical stimulation, sensory drive, and stochastic noise—enter every level of the hierarchy through a unified forcing decomposition. By grounding diverse NMM formalisms in a single dynamical core, the framework provides both experimentalists and theorists with a navigable common language for model comparison, parameter mapping, and the design of brain stimulation protocols across scales.

Every neural mass model, from the simplest harmonic oscillator to biologically derived mean-field equations, is secretly the same push–pull dance between excitation and inhibition.

Computational neuroscience has a model zoo problem. Researchers studying brain oscillations reach for different formalisms — Kuramoto, Wilson–Cowan, Jansen–Rit, Stuart–Landau — depending on their subfield, and these models rarely talk to each other. Translating results across them feels arbitrary, because nobody has laid out the explicit mathematical steps connecting them. This paper does exactly that.

The central observation is elegant: every neural mass model, regardless of complexity, contains a pair of variables in antagonistic coupling. One pushes the system away from rest; the other pulls it back. In the simplest case this is just ẋ = −ωy, ẏ = ωx — two variables chasing each other in circles. The paper shows this "push–pull motif" survives intact as you climb a six-rung ladder of increasing biological realism: from the undamped harmonic oscillator, through the Stuart–Landau (which adds a cubic nonlinearity to stabilize amplitude via a Hopf bifurcation), to Wilson–Cowan (where the abstract variables become concrete excitatory and inhibitory firing rates with sigmoidal transfer functions), to the Jansen–Rit family (which replaces instantaneous synaptic responses with realistic second-order filters), and finally to next-generation mean-field models that are derived exactly from populations of spiking neurons — no phenomenological assumptions required.

What makes this more than a taxonomy is that the paper provides explicit mathematical transformations between rungs. Each step up the ladder introduces one new ingredient — nonlinearity, synaptic filtering, a dynamic transfer function — and the paper shows precisely what that buys you and what it costs. This lets you map parameters and results across formalisms in a principled way rather than by intuition. The paper also standardizes how external inputs enter every model through a unified decomposition: physiological drive plus electric-field coupling (relevant for brain stimulation like tES or TMS) plus noise. That means a stimulation protocol designed at one level of the hierarchy can be translated to predictions at another.

The practical payoff is a decision tree for model selection. Studying synchronization across a large network? Use Kuramoto — it's analytically tractable. Modeling pharmacological E–I balance? Wilson–Cowan. Fitting EEG spectra with realistic alpha-band dynamics? Jansen–Rit. Need a rigorous link from single-neuron spiking to population activity? The QIF-derived next-generation model. The source is written as a blog post accompanying an arXiv preprint, so it is more expository than a full methods paper, but the equations are present and the conceptual scaffolding is complete.