A Control-Theoretic Formulation of Global Workspace Theory

What the math gives to ML

The paper provides a control-theoretic way to distinguish a genuine internal mediator from a component that merely receives or broadcasts information. Its transferable asset is the joint use of finite-horizon reachability, observability, and a boundary Hankel operator, which identifies internal modes that are both excitable from the surrounding network and able to affect it again. This suggests a trainable routing regularizer for neural modules: reward candidate blocks whose internal states have broad, aligned input-output pathways, while penalizing one-sided hubs and split read/write pathways. The main uncertainty is computational cost and whether the resulting dynamical organization improves task performance rather than merely matching a control metric.