A modular state-space model of human perception, cognition, and decision dynamics

What the math gives to ML

The paper provides a transferable modular state-space construction in which perception is a contractive latent estimator and cognition evolves as a bounded, input-driven discrete dynamical system. Its most useful asset for neural networks is not the psychological interpretation, but the sufficient conditions for forward invariance, contraction of perceptual inference, and input-to-state stability of the cognitive state. These conditions can be converted into a recurrent or state-space architecture with separately constrained perception and cognition blocks, together with a computable step-size or spectral-gain ceiling. The resulting model should exhibit bounded hidden states and exponentially decaying sensitivity to initial conditions below a measurable contraction boundary.