Perceptual inference (interoceptive)
Named and elaborated in Farb et al. (2015) as the underspecified second route to prediction-error (PE) minimization within Seth’s interoceptive-inference framework — already implicit in predictive-coding generally (“updating generative models”) but here given a full clinical/contemplative treatment as the counterpart to active-inference.
Mechanics
Where active inference reduces PE by shaping sensation to match a (high-precision) prior, perceptual inference reduces PE by updating the prior to match sensation — increasing the precision weighting of sensory information relative to priors. The posterior probability updates the prior rather than prompting action on the world. A low-specificity range of expected body states means fewer sensations register as surprising, lending future priors (i.e., updated simulations) greater accuracy relative to actual bodily variability.
The contemplative mapping
Modern psychological emotion regulation (suppression, distraction, reappraisal) is characterized as active inference — techniques that alter the interoceptive signal or attention to it. Contemplative techniques (acceptance, equanimity, continuous non-interfering observation) are characterized as perceptual inference — changing one’s relationship to sensation rather than the sensation itself. Appraisals still follow perception under this account, but become less rigidly stereotyped.
The clinical evidence underneath the mapping
The active-vs-perceptual contrast is not only theoretical: Farb et al. (2011) measured its neural twin with a hard outcome attached, four years before the 2015 framework named it. In remitted depressed patients, an elaborative/self-referential (mPFC) mode of reacting to a sad film — the maladaptive counterpart of active-inference regulation, tracking rumination — predicted relapse, while a sensory/accepting mode (tracking acceptance, the decentered stance closest to perceptual inference) predicted sustained remission. It is the wiki’s one prospective demonstration that staying perceptual, rather than elaborating, has clinical value — with the standing caveat that the sensory pole measured there was exteroceptive (visual), so its interoceptive application is Farb’s later generalization. See cognitive-reactivity.
The earlier Farb et al. (2010) paper closes part of that caveat from the other direction. There the sensory pole under emotional challenge is the insula — genuinely interoceptive — and mindfulness-trained participants recruited it more (kept it from deactivating) while elaborating less, with insula recruitment tracking lower depression. Read through this framework, that is the interoceptive perceptual-inference mode caught in the act: staying with the felt body rather than shaping or narrating it. What 2010 lacks that 2011 has is prospectiveness — it is a cross-sectional between-groups association, not a forecast — so the two papers trade region for design (see insular-cortex).
Self-flagged controversy
The authors explicitly note: “We note that this conceptualization of acceptance or equanimity as a ‘bottom–up’ perceptual rather than ‘top–down’ cognitive strategy may be controversial.” This matters because both active and perceptual inference are, in the predictive-coding architecture, top-down processes (both involve priors and precision-weighting) — so labeling perceptual inference “bottom-up” is a looser, functional usage (attention shifts toward the sensory end of the hierarchy) rather than a claim about feed-forward processing per se. This is a narrower and distinct question from the feedforward-vs-predictive-interoception debate about whether interoception overall is predictive or feed-forward.
Untested hypothesis
Farb et al. hypothesize that because most people’s bodies show more variation than they expect, contemplative practice’s effect is to broaden (de-precisify) the distribution of priors rather than narrow it around any single expected experience — yielding a simulation map that more accurately tracks actual sensory experience.
Clinical stakes
Moving directly to active inference (regulation) without perceptual inference (exploration) forecloses insight — e.g., “knee-jerk” regulatory responses to arousal may drive emotional eating without ever testing whether the arousal’s inferred cause (hunger) was correct. Both routes are needed; healthy self-regulation is framed as iterative cycling between them, not the primacy of one.
Open question: is perceptual inference genuinely a regulatory mode?
A sharper version of the self-flagged controversy above: is perceptual inference a distinct regulatory strategy at all, or does the very act of choosing to “let go” — redirecting attention, lowering the precision assigned to one’s own priors, permitting an update — already have the shape of an active-inference act? If so, perceptual inference may never occur as a genuinely unchosen event; it could always be downstream of a prior, meta-level active-inference move. See perceptual-inference-as-regulation for the full framing — this is flagged there as a live, unresolved question at the center of current thinking on this material, not a settled matter.