Karl J. Friston

The theorist behind the predictive-coding / free-energy apparatus this wiki runs on. Cited by name across the interoception literature here — Seth, Barrett, Stephan, Petzschner, Quadt, Harrison all lean on his framework — but held entirely secondhand until “A theory of cortical responses” (2005) was ingested. That paper is his statement of the machinery; this page is the person behind it.

Central idea: one principle for perception, learning and action

Friston’s claim is that a self-organizing system that resists disorder must minimize the free energy of its sensory states — a statistical-physics bound on surprise (negative log model evidence). From this single imperative he derives:

• Perceptual inference — updating internal representations to reduce prediction error (the E-step; perceptual-inference).
• Perceptual learning — updating the generative model’s parameters, i.e. synaptic plasticity, over slower timescales (the M-step).
• active-inference — changing sensory input by acting on the world, so that descending predictions are fulfilled rather than revised.

The interoceptive literature in this wiki uses all three, usually without noticing they are one optimization.

Why he matters here without being an interoception researcher

Friston works on cortical dynamics, neuroimaging methods (he originated statistical parametric mapping and dynamic causal modelling), and the general theory of the Bayesian brain. Interoception enters his work chiefly through the collaboration with anil-seth — Seth & Friston (2016) — where the free-energy formalism is applied to the internal milieu and the agranular visceromotor cortex argument is made. Everywhere else in this wiki he is the upstream source: the predictive-coding page’s “core mechanics,” the cost-of-prediction-error Barrett builds on, Stephan’s allostatic-self-efficacy account, and the computational-psychiatry programme all presuppose his scheme.

The bet that is still open

His most consequential and least-settled commitment is anatomical: that prediction error and representation are carried by distinct, explicitly represented neuronal populations, hierarchically arranged. It is elegant and motivated by cortical microcircuitry, but it is a hypothesis. The wiki’s Petzschner et al. (2021) entry places predictive coding as one algorithmic proposal among several for the computational-level claim of Bayesian inference; Harrison et al. (2021) is the first human interoceptive test and did not find the prediction-vs-error dissociation his architecture predicts. See feedforward-vs-predictive-interoception.