Deep Equilibria: Existence and Computability

Abstract

We introduce a general concept of layered computation model, of which neural networks are a particular example, and combine tools of topological dynamics and model theory to study asymptotics of such models. We prove that, as the number of layers of a computation grows, the computation reaches a state of ``deep equilibrium" which amounts to a single, self-referential layer. After proving the existence of deep equilibria under fairly general hypotheses, we characterize their computability.