Active inference
The second route to prediction-error minimization: rather than changing the model to fit the world, the agent changes the world to fit the model. In predictive-coding accounts of motor control, descending proprioceptive predictions are fulfilled by classical motor reflexes (Adams, Friston et al.).
Seth’s interoceptive application
In Seth (2013), active interoceptive inference is the mechanism by which interoceptive predictions are “transcribed into reference points for autonomic reflexes that regulate physiological homeostasis” — sympathetic/parasympathetic outflow from AIC/ACC enslaves autonomic reflexes (heart rate, respiration, smooth muscle) just as proprioceptive predictions enslave motor reflexes.
The precision precondition
Active inference works only if the relevant prediction errors have transiently low precision (attenuated attention). Otherwise precise interoceptive prediction errors would revise the generative model rather than drive autonomic action. This gives a principled account of why phasic physiological changes accompany, e.g., rubber-hand-illusion induction (threat to the rubber hand → skin-conductance responses), read as altered autonomic control under descending self-related predictions. See experience-of-body-ownership.
Farb et al.’s regulatory pairing
Farb et al. (2015) pair active inference with its named counterpart, perceptual-inference (weighting sensation over priors instead), and map modern psychological regulation strategies — suppression, distraction, reappraisal — onto active inference, contrasted with contemplative strategies (acceptance, equanimity) mapped onto perceptual inference. They caution that both routes are needed; over-reliance on active inference can foreclose the insight that perceptual inference affords (e.g., “knee-jerk” regulatory responses in emotional eating).
Does perceptual inference reduce to active inference?
An open question worth holding against the pairing above: if shifting into perceptual inference must itself be initiated (attention redirected, precision on one’s own priors lowered, an update permitted), that initiating move looks like active inference — suggesting the two are not peers but nested, with active inference operating at the meta level even when the resulting content is “letting go.” See perceptual-inference-as-regulation.
It has a rival, and they are the same theory at one level (Petzschner et al. 2021)
Petzschner et al. (2021) set interoceptive active inference (IAI) beside homeostatic reinforcement learning (HRL) — the first time the wiki has an alternative account of how the body gets regulated that is not a variant of the free-energy principle.
Their IAI summary is the cleanest statement of it in the wiki. During interoception, prediction errors update the model and infer the state. IAI’s addition is that the predictions themselves represent a desired future internal state, so instead of updating the prediction, the organism can act to fulfil it — “this corresponds to reversing the inverse model, turning it into a forward model again.” Both routes reduce prediction error, hence minimize surprise. The rationale is that priors encode the states an organism is likely (and needs) to occupy, so fulfilling them maintains homeostasis and allostasis automatically: “Under active inference, agents stay alive by predicting the states that keep them alive, and act to fulfill those predictions” (Morville et al.).
Then the verdict, in two parts:
Same, computationally. “The drive in HRL can be formally re-expressed as surprise in IAI” (Hulme et al. 2019). Both explain the same repertoire beyond reflex — context-adjusted reflexes, anticipatory responses, Pavlovian/habitual/goal-directed control. The free-energy vocabulary is not, at this level, doing distinctive explanatory work; a different tradition arrives at the same place.
Different, implementationally — and this is the testable part. IAI requires predictions and prediction errors to be explicitly represented in neuronal populations, hierarchically arranged, topped by visceromotor cortex projecting down to hypothalamus, PAG and parabrachial nucleus, where descending predictions set internal set-points directly and can activate or suppress low-level reflex arcs. (Supporting evidence offered: Levinthal & Strick’s finding that rostral insula, mPFC and primary motor cortex influence sympathetic and parasympathetic output to the stomach.) HRL requires none of that — and could run without any interocept at all.
There is also an architectural fork about how multiple controllers coexist: in HRL, reflexive/Pavlovian/instrumental controllers compete or collaborate; in IAI they are integrated as layers of a single hierarchy. The wiki should note that Farb et al.’s active/perceptual-inference pairing and the perceptual-inference-as-regulation debate are arguments within the second picture — they presuppose the hierarchy that HRL declines to posit.
Attention is covert action (Parr et al. 2019)
Seth & Friston’s epistemic mode above is usually pictured as overt — a saccade, an exploratory movement. Parr, Corcoran, Friston & Hohwy (2019) make the sharper point that the same epistemic action can be entirely covert: attention is itself a policy, selected to minimise expected free energy, that sets the precision of a likelihood mapping (“how uncertain am I about what I would see if I attended there?”). This collapses two things the wiki otherwise states separately — attention-as-action and attention-as-precision-weighting — into one object.
They demonstrate it by simulating two perceptual phenomena. Troxler fading (a fixated peripheral percept dissolving) is the accumulation of uncertainty about an unsampled stimulus, once the volatility prior in the generative model is allowed to relax the posterior toward maximum entropy — a percept fades because the belief about it has lost its information, not because a receptor fatigued. Binocular rivalry is the same, with a covert attentional action doing the sampling: the percept alternates because attention alternates to whichever hypothesis carried the most resolvable uncertainty last, with no overt movement at all. The paper is exteroceptive throughout, but it is the formal statement of the identity Farb (2013) shows empirically in the body — that attention decides which cortex represents a given interoceptive signal — and of the aberrant-precision individual differences the wiki invokes for interoceptive psychopathology (see interoceptive-precision).
Two kinds of active inference (Seth & Friston 2016)
Seth & Friston (2016) split active inference itself into two modes, both still active (action-based): epistemic active inference selects actions expected to increase the fit between predictive models and hidden causes (saccades, exploratory movement — information-seeking, not regulatory), while instrumental active inference leverages an existing model to control sensory variables (autonomic/interoceptive regulation; Powers’s perceptual control theory — “control systems control what they sense, not what they do”). Notably, the paper’s own taxonomy of active inference has no third, passive/perceptual mode alongside these two — directly relevant to whether Farb et al.’s perceptual-inference is a genuine peer category or, per perceptual-inference-as-regulation, always nested under some initiating active-inference act.