Bridging Neurons to Behaviour: A Generative Neural Engine Mechanistically Rejects the Independent Race Model

Abstract

The brain produces robust, low-dimensional behaviour from variable, high-dimensional activity. The dominant account of action control, the Independent Race Model (IRM), explains stopping as a "Go" and a "Stop" process racing independently, and predicts behaviour accurately. But independence is a behavioural assumption never confronted with neural dynamics, and the same premotor neurons drive both processes. We therefore trained a generative Deep Markov Model on premotor activity from two macaques; trained on neural data alone, it emergently reproduces the full reaction-time distribution, validating it as a proxy for the neural hardware. Used for in silico experiments, this engine reveals systematic violations of both of the IRMs independence axioms, with violations emerging from the geometry of a single shared manifold. Apparent independence was never real: an artifact of observing behaviour without seeing the manifold that produces it.