Active interoceptive inference and the emotional brain

Study at a glance

• Authors: Anil K. Seth, Karl J. Friston
• Year: 2016
• Journal: Philosophical Transactions of the Royal Society B
• DOI: 10.1098/rstb.2016.0007
• Design: theoretical/review (predictive-coding and free-energy formalism)

Key findings

• Restates interoceptive inference with Friston’s free-energy/active-inference formalism: autonomic reflexes are ‘enslaved’ by descending predictions from deep generative models of the internal and external milieu — bodily states are regulated, not merely represented.
• Distinguishes epistemic active inference (actions selected to increase the fit between models and hidden causes — e.g., saccades, exploratory body movements) from instrumental active inference (actions that leverage predictive models to achieve control of sensory variables, per Powers’s perceptual control theory: ‘control systems control what they sense, not what they do’).
• Proposes visceromotor areas (VMAs) — anterior insular cortex (AIC), anterior cingulate cortex (ACC), subgenual cortex (SGC), and orbitofrontal cortex (OFC) — sit atop the interoceptive hierarchy, embody a generative model of interoceptive responses, and issue descending predictions that serve as homeostatic set-points.
• Grounds this hierarchy anatomically: VMAs are agranular/dysgranular, lacking a well-formed granular layer IV — the cytoarchitectural signature of cortex specialized for issuing top-down predictions rather than receiving bottom-up input, giving the predictive architecture a structural (not just functional) basis.
• VMAs project descending connections to periaqueductal grey (PAG) and parabrachial nucleus (PBN) for visceromotor control, and receive ascending viscerosensory input from posterior/mid-insula; visceromotor efferents also directly innervate viscerosensory areas, potentially providing efference copy/corollary discharge.
• Reframes the homeostasis/allostasis balance as a precision-weighting problem: confidence (precision) in deeper allostatic expectations determines whether low-level homeostatic reflexes (e.g., mobilizing glucose stores) or goal-directed allostatic behaviour (e.g., preparing and eating a meal) predominates.
• Extends the framework to autism: proposes a failure to attenuate the precision of interoceptive prediction errors, producing autonomic hypersensitivity and an inability to ignore the absence of expected interoceptive cues during affiliative interaction (e.g., breastfeeding) — with implications for self/other differentiation and attachment. Introduces the ‘interoceptive trait prediction error’ (ITPE): the mismatch between objective interoceptive accuracy (heartbeat detection) and subjective interoceptive sensibility (questionnaire), elevated in autism and correlated with anxiety across both autistic and control samples.
• Extends the framework to depression and fatigue: chronic dyshomeostasis from noisy/imprecise interoceptive afferents (endocrine/immunological) increases reliance on increasingly unreliable interoceptive priors, tipping into fatigue and sickness behaviour; alternatively, fatigue and depression are metacognitive beliefs about low ‘allostatic self-efficacy’ following prolonged dyshomeostasis, echoing learned-helplessness accounts.
• Extends embodied selfhood into a graded hierarchy from ‘low’ (being/having a body) through first-person perspective and agency to ‘high’ (narrative self, social self) — interoception structures the low end and may shape the higher levels too. Cites cardio-visual and respiratory-visual feedback as evidence interoception shapes body-ownership illusions (rubber hand, full-body illusion).
• Concludes with an inversion of Descartes’s ‘beast machine’ doctrine: subjective selfhood may arise because of, not in spite of, the fact that we are physiologically regulated organisms.

Limitations

• Explicitly states direct empirical evidence for interoceptive predictions or prediction errors (as distinct signals) is ‘still lacking’ — the case rests on anatomical plausibility, circumstantial functional evidence, and analogy to exteroceptive/motor predictive coding.
• The autism and depression/fatigue applications are labeled speculative; the two competing depression accounts (noisy afferents vs allostatic self-efficacy beliefs) are not adjudicated between.
• A theoretical/review piece, not a new empirical test; proposes laminar fMRI and heartbeat-variability analyses as future tests but does not conduct them.

Connections

• Methods: Heartbeat detection task, Rubber hand illusion (RHI)
• Concepts: Interoceptive inference, Predictive coding, Active inference, Allostasis, Homeostasis (and feelings as homeostatic sensations), Visceromotor areas (VMAs), Interoceptive trait prediction error (ITPE), Embodied selfhood, Insular cortex, Anterior cingulate cortex (ACC), Salience network
• Debates: inferential?, Is perceptual inference itself a genuine regulatory mode, or does the act of ‘choosing to let go’ collapse it back into active inference?
• Related studies: Seth (2013), Craig (2009), Lindquist et al. (2012)

The formal successor to Seth (2013), co-authored with Karl Friston, published in a Royal Society theme issue on “Interoception beyond homeostasis.” It restates Interoceptive inference using the full free-energy/active-inference apparatus and adds two things the 2013 paper lacked: a structural (cytoarchitectural) argument for why the interoceptive hierarchy should be read as predictive rather than feed-forward, and worked clinical extensions (autism, depression/fatigue).

Epistemic vs instrumental active inference

A distinction largely absent from Seth (2013): epistemic active inference selects actions expected to increase the fit between predictive models and the hidden causes of sensation (saccades; exploratory movements that inform self-models) — a curiosity-driven, information-seeking mode. Instrumental active inference leverages an existing predictive model to achieve control of sensory variables, invoking Powers’s perceptual control theory: “control systems control what they sense, not what they do.” Interoceptive/autonomic regulation is instrumental active inference — descending predictions set homeostatic points that autonomic reflexes fulfill. Both are still active (action-based) routes to prediction-error reduction; the paper does not posit a passive/perceptual third mode alongside them, which is directly relevant to Is perceptual inference itself a genuine regulatory mode, or does the act of ‘choosing to let go’ collapse it back into active inference? — see that debate page for the tension this raises against Farb et al.’s active/perceptual dichotomy.

Visceromotor areas: an anatomical argument for prediction

Where Seth (2013) placed the AIC alone as comparator, this paper broadens and sharpens the neuroanatomy: visceromotor areas (VMAs) — AIC, ACC, subgenual cortex (SGC), and orbitofrontal cortex (OFC) — sit at the top of the interoceptive hierarchy. The key new argument is cytoarchitectural: these regions are agranular or dysgranular, lacking a well-formed granular layer IV (the canonical target of feedforward, prediction-error-carrying input in laminar predictive-coding models). Structure, not just function, marks them as a source of descending predictions rather than a recipient of ascending error. This gives Seth’s side of inferential? a harder anatomical foothold than the functional/circumstantial evidence Seth (2013) relied on.

VMAs receive ascending viscerosensory input from posterior/mid-insula and send descending projections to PAG and the parabrachial nucleus (PBN) for visceromotor control; visceromotor efferents also directly innervate viscerosensory areas, potentially supplying efference copy/corollary discharge that helps construct ascending interoceptive prediction errors.

Homeostasis and allostasis as a precision problem

Refines Allostasis: whether the system falls back on low-level homeostatic reflexes (e.g., hypoglycemia triggering glucose mobilization) or engages higher, goal-directed allostatic behaviour (preparing and eating a meal) is cast as a function of the confidence (precision) placed in deeper expectations about future action — not a qualitatively different system, but a different point on the same precision-weighted hierarchy.

Clinical extensions

• Autism: a failure to attenuate the precision of interoceptive prediction errors produces autonomic hypersensitivity and precludes learning that a nurturing (m)other is the same object across affiliative interactions (since the absence of expected interoceptive cues, e.g. during breastfeeding, cannot be discounted) — with knock-on effects for self/other differentiation, attachment, and central coherence. Introduces the interoceptive trait prediction error (ITPE), operationalized as the mismatch between objective heartbeat-detection accuracy and subjective interoceptive-sensibility questionnaires; elevated in autism and correlated with self-reported anxiety in both autistic and control samples (recalling Paulus & Stein’s non-Bayesian anxiety-as-interoceptive-prediction-error proposal). Flags a confound: autism’s frequent co-occurrence with alexithymia, and inconsistent findings across studies on the direction of interoceptive-sensibility change in autism.
• Depression and fatigue: two (not-yet-adjudicated) accounts. (1) Noisy/imprecise interoceptive afferents (endocrine, immunological) reduce precision-weighting of ascending interoceptive signals, increasing reliance on now-dysfunctional interoceptive priors — a positive-feedback spiral into dyshomeostasis, fatigue, and sickness behaviour. (2) Fatigue and depression as metacognitive beliefs about low allostatic self-efficacy following prolonged experienced dyshomeostasis, paralleling learned-helplessness accounts. Both are linked to agranular visceromotor cortex pathophysiology in depression.

Relation to Seth (2013)

Largely the same core claim (emotion as interoceptive predictive coding; AIC/VMAs as comparator; active inference as autonomic enslavement), restated with Friston as co-author and with three substantive additions: the agranular-cortex structural argument, the epistemic/instrumental active-inference distinction, and the autism/depression clinical program. Read together with Craig (2009) and Lindquist et al. (2012), the AIC/ACC/OFC complex this paper calls “VMAs” is very close to the “core affect” network Lindquist’s meta-analysis identifies empirically — a convergence between a predictive-coding neuroanatomical argument and a psychological-constructionist meta-analytic one, made independently and from different starting assumptions. See Visceromotor areas (VMAs).