Predictive Processing & The Free-Energy Principle

Hermann von Helmholtz, in Handbuch der physiologischen Optik (1867), proposed that perception is unconscious inference: the brain infers the most-probable cause of its sensory inputs given prior expectations. The idea was radical and largely forgotten across the next century, when behaviorism and information-processing cognitive science organized psychology around stimulus-response framings. Karl Friston at UCL synthesized it into a coherent neuroscientific programme starting around 2005. The brain, on this account, is not primarily a stimulus-response device but a prediction machine that generates top-down expectations about sensory inputs, compares them to actual data, and either updates its model (perception) or changes the world (action) to reduce the mismatch.

Predictive processing is the empirically tractable level. Cortical hierarchies generate predictions about the activity of layers below; prediction errors propagate upward; the system updates predictions to minimize error. The Rao-Ballard 1999 model gave the canonical implementation. It accounts for a striking range of phenomena: end-stopping, bistable perception (the Necker cube flipping between competing hypotheses), Adelson's checker-shadow illusion, Kanizsa-triangle illusory contours. Anil Seth's controlled hallucination framing — perception as the brain's best guess at what's out there, with sensory input correcting rather than constructing experience — is the accessible form of the view. Action in this framework is active inference: the agent acts on the world to make sensory data conform to its predictions, with motor neurons firing because they correct prediction error about expected proprioception. The framework dissolves the perception-action divide. The free-energy principle generalizes the picture: any system maintaining its boundaries against entropic dissipation is doing something equivalent to free-energy minimization. The mathematics — variational free energy, Markov blankets, active inference under generative models — is non-trivial. Empirical predictions include specific cortical activity patterns, precision-weighting in attention, and psychiatric disruptions (autism as over-precise prediction error; schizophrenia as miscalibrated precision producing hallucinations from internal predictions). Whether the framework is a substantive theory or a tautology that fits anything is contested.