Open-source neurodynamical computing

White-box AI with explainable neurodynamics

An alternative to black-box LLMs using established neurodynamical models. CPU-first, lightweight, and fully explainable.

Why CortexAI?

Every output can be traced back to the equations that produced it.

Lightweight, low memory footprint, low latency. No GPU required.

Each equation is an independent module, composable in a graph.

All performance metrics are reproducible and publicly benchmarked.

Each module implements a well-established mathematical model from computational neuroscience

Synchronisation de phase par couplage non-linéaire

d\phi_i/dt = \omega_i + \frac{K}{N} \sum_{j=1}^{N} \sin(\phi_j - \phi_i)

Kuramoto, 1975

Dynamique de populations excitatrices/inhibitrices

\tau_E \frac{dE}{dt} = -E + S(W_{EE} \cdot E + W_{EI} \cdot I + I_{ext})

Wilson & Cowan, 1972

Modèle de neurones à spikes (compromis biologique/coût)

\frac{dv}{dt} = 0.04v^2 + 5v + 140 - u + I

Izhikevich, 2003

Analyse de stabilité par exposants de Lyapunov

\lambda = \lim_{t \to \infty} \frac{1}{t} \ln \frac{|\delta Z(t)|}{|\delta Z_0|}

Lyapunov, 1892

Signal d'erreur de prédiction de récompense

\delta = R(t) - V(t)

Schultz, 1997

<88ms

Latency (100 oscillators)

<231MB

Memory (1000 neurons)

CPU

No GPU required