The sensory-control loop

From Petzschner et al. (2021). Adaptive behaviour is circular, not linear: internal states produce sensory signals; the CNS turns those signals into actions; the actions change the states, and therefore the next signals. Three arcs, and each is a place a computational model can live.

arc	the model answers	frameworks in the wiki
states → sensations (models of interoception)	how is a hidden internal state inferred from noisy afferents?	Bayesian inference; predictive-coding / interoceptive-inference
sensations → action (models of body regulation)	which regulatory action, in this context?	active-inference; homeostatic-reinforcement-learning
action → states (forecasting)	how will internal states evolve, under this action and under their own dynamics?	almost nothing

The framing is deliberately agnostic. It does not say which model is right; it says what each model is a model of — which turns out to be the fastest way to see that the wiki’s sources pile onto the first two arcs and leave the third nearly empty.

Why the loop, and not a set-point

The degenerate case of the loop is the reflex arc: a comparator holds a fixed set-point, deviation triggers a hard-wired response. The baroreflex is the worked example — baroreceptors signal a pressure rise, NTS efferents activate parasympathetic outflow and inhibit vasopressin release, pressure falls. Adaptive, and limited: in a changing environment the default reaction is sometimes the wrong one.

Flexible control is defined against that baseline in two directions, and the distinction is cleaner than the wiki’s usual homeostasis/allostasis pairing:

• Predictive homeostasis — temporarily moving away from the set-point in anticipation of a perturbation (raising body temperature before an expected cold challenge).
• allostasis — changing the set-points themselves.

Both are documented across species: homeostatic control adjusts to context, to expected future events, and — via placebo effects (Wager & Atlas) — even to abstract beliefs. That last one is the strongest empirical warrant the paper offers for the whole inferential turn: a set-point that moves with a belief is not a thermostat.

The internal model, and Conant & Ashby

The loop as drawn is the world. The claim about the brain is that the CNS represents it — an internal model, in the cybernetic sense captured by the principle that “every good regulator of a system must be a model of that system” (Conant & Ross Ashby, 1970). The wiki already holds that principle: it is step 1 of Barrett’s derivation on allostasis, where she uses it to get from body-budgeting to interoception. Petzschner et al. use the same premise more modestly, as the licence to talk about internal models at all.

The three arcs then become three internal models — an inverse model running states-from-sensations (that is interoception), a forward model running actions-from-states (that is body regulation), and a forward model running states-from-actions (that is forecasting). The IAI trick, in this vocabulary, is to reverse the inverse model back into a forward model: treat the prediction as the desired state and act to fulfil it.

Where the wiki’s material sits on it

• Craig describes the afferent limb — the first arc’s anatomy, with no model of the arc.
• Seth, Seth & Friston and Barrett model arcs one and two together, in one formalism.
• Keramati & Gutkin’s HRL models arc two without necessarily modelling arc one at all.
• Nobody in raw/ models arc three. forecasting records the gap.

What the device is good for, and its limit

Its value is that it separates what a framework explains from what vocabulary it explains it in, which is what lets the paper report that HRL and interoceptive active inference are equivalent at the computational level while differing at the implementational one. Its limit is the same thing: it is a taxonomy, not a claim, and it cannot tell you which model is true. What it can tell you is when two apparently rival theories are competing for the same slot — and when, as with the forecasting arc, nobody is competing for a slot at all.