Computational Models of Interoception and Body Regulation

Petzschner, Garfinkel, Paulus, Koch & Khalsa (2021) — Computational Models of Interoception and Body Regulation

The wiki’s first source that treats the computational account of interoception as a field with alternatives rather than as the theory. Everything the wiki holds under predictive-coding, interoceptive-inference and active-inference comes from Seth, Friston, Barrett and Khalsa — four sources in one lineage. This review, from the roadmap’s own author cluster (Khalsa, Paulus, Garfinkel) plus Petzschner and Christof Koch, steps back far enough to show that lineage occupying one branch of a larger tree, and names what the other branches are.

The first article in the Trends in Neurosciences January 2021 special issue on interoception. Of the issue’s other reviews, Bonaz et al. on diseases and disorders has since been read — it cites this paper as its reference 118, for the computational modelling of depression and anxiety — and Quigley et al. on functions and Weng et al. on interventions have followed, leaving Berntson & Khalsa on neural circuits and the Furman editorial unread. Weng et al. cite this paper (their reference 5) in their closing suggestion that interoceptive interventions “be integrated with computational modeling… to understand how interoceptive representations may be altered in specific clinical conditions” — and thank Petzschner for consultation on the manuscript, so the computational framing in the intervention review is partly this one’s. The two read so far divide cleanly: this is the computational quadrant of the special issue, Bonaz et al. the clinical-medicine one, and they agree on the item that matters most for the wiki’s translational claims — that the biomarker cupboard is empty. This paper: “there is little empirical work testing such models’ predictions.” Bonaz et al., of the markers actually in research use: “few of these measures have yet found broad clinical application in this role.”

The organizing device: the sensory-control loop

The paper’s real contribution is a piece of bookkeeping that turns out to be clarifying. Adaptive behaviour is a closed loop — internal states produce sensory signals, signals inform action selection, actions change the states — and every computational model of interoception is a model of one arc of that loop:

arc	what it models	frameworks
states → sensations	how the internal state is inferred from noisy afferents	interoceptive Bayesian inference; interoceptive predictive-coding
sensations → action	how the right regulatory action is selected	homeostatic-reinforcement-learning; interoceptive active-inference
action → states	how internal states will evolve	forecasting — barely formalized

See sensory-control-loop. The device does two useful things: it makes visible that the wiki’s sources cluster on the first two arcs and almost nothing addresses the third, and it separates the reflex arc (a degenerate loop with a fixed set-point) from the flexible control the field is actually trying to explain — predictive homeostasis (temporarily leaving the set-point) and allostasis (moving the set-point).

Levels, and a correction to how the wiki has been reading Seth

The paper is careful in a way the wiki’s other predictive sources are not, and it changes the status of a claim recorded across several pages here. Bayesian inference is a computational-level (Marr) statement: it says what an ideal agent would compute, and functions as a benchmark against which real behaviour is compared. It “does not come with a prescription for how the computations are implemented.” Predictive coding is one algorithmic proposal for implementing it — “one of the most prominent,” not the only one, and the paper cites alternatives (probabilistic population codes, sparse coding) and one outright dissent (Brette, “Is coding a relevant metaphor for the brain?”).

So: when the wiki records that interoception “is” predictive coding, it is running two claims together. The Bayesian claim (the brain combines noisy afferent likelihood with a prior, weighted by precision) is far better supported than the specific claim that it does so via hierarchical prediction-error units. The paper’s own status report on the latter is blunt: “To date, the full interoceptive brain network underlying this implementation has not been identified.” This sharpens rather than contradicts what feedforward-vs-predictive-interoception already records.

One definitional consequence worth keeping. On this framing, “interoception and exteroception can thus be distinguished based on the type of state being inferred — internal or external — rather than the specific sensory channels that contributed to the inference.” Vision counts as interoceptive input when it informs an internal state. That is a cleaner criterion than the wiki’s usual channel-based one, and it is why Farb (2011)‘s exteroceptive sensory pole is less anomalous than the wiki has been treating it.

The other quiet reframing: priors replace set-points. In the reflex arc there is a set-point and a comparator; in the inferential account, the states an organism expects to occupy are the prior, and they coincide with survivable states because evolution and experience put them there. Whether those priors are innate or learned is left open, and flagged as a question for developmental work — see social-vs-biological-origins-of-interoception, where the wiki’s version of that question already sits.

The rival the wiki did not have: homeostatic reinforcement learning

homeostatic-reinforcement-learning is the genuinely new material here. Keramati & Gutkin’s proposal redefines reward homeostatically — an outcome is rewarding to the extent it reduces drive, the distance between the current and desired internal state — so ordinary reinforcement learning machinery (value estimates, reward prediction errors, dopaminergic midbrain) delivers homeostatic control without any inference about bodily states at all.

Two things make this important for the wiki:

1. It is formally equivalent to active inference, at one level. “The drive in HRL can be formally re-expressed as surprise in IAI.” Both explain the same extensions beyond reflex (context-adjusted reflexes, anticipatory responses, Pavlovian/habitual/goal-directed control). So the free-energy framing the wiki has been treating as the deep account of why organisms act is, at the computational level, one notation for a result the reinforcement-learning tradition reaches independently. That is a strong statement in favour of the content of the account and a weak one in favour of its vocabulary.

2. It comes apart from active inference at the implementational level — and that is testable. IAI needs explicit prediction and error populations, arranged hierarchically, with visceromotor cortex at the top projecting to hypothalamus, PAG and parabrachial nucleus. HRL does not: internal states can be signalled directly into reward computation (orexin neurons projecting lateral hypothalamus → VTA; ghrelin, leptin and insulin receptors in VTA; expected drive reduction via the opioid system). The paper draws the consequence explicitly — this “rais[es] the possibility that this type of control could be performed in the absence of an interocept.”

That sentence is the most consequential thing in the paper for this wiki. Nearly every page here assumes that regulating the body requires representing it. HRL says a large part of body regulation may not, and reserves inference for the complex goal-directed cases. It is the same architectural question the somatic-feedback debate asks about decisions (body loop vs as-if loop; whether the signal must be felt), arriving from control theory instead of neuropsychology.

The two frameworks also differ in how they get multiple controllers to coexist: HRL has reflexive, Pavlovian and instrumental controllers competing or collaborating; IAI has them integrated as layers of one hierarchy. That is a real architectural fork, and the paper’s honest verdict is that the computational level cannot decide it — “these implementational differences may serve as the basis for future research programs.”

What models of interoception have to handle and mostly don’t

The paper’s most concrete service is a list of ways internal signals are unlike the visual signals predictive coding was built on. The wiki’s methods pages have been circling several of these without stating them as a class:

• Sparse, distributed transduction. Internal organs are sparsely innervated relative to exteroceptive ones, and a single event is split across sensors — a heartbeat is encoded separately as the occurrence of a pulsation, its strength, the associated pressures, and the chemical consequences. “Heartbeats are thus transduced in a distributed manner quite distinct from vision or audition.” This is a mechanistic gloss on why heartbeat-counting performance is so strange: there is no single “heartbeat signal” to be accurate about. See is-the-heartbeat-counting-task-valid.
• Timescales spanning five orders of magnitude. Cardiac and respiratory changes in seconds; gastric and osmotic in minutes to hours; immune in minutes to weeks (but anaphylaxis in seconds). Models assuming fast stimulus-response associative learning cannot span this. The wiki’s cardiac-heavy evidence base is a methodological consequence of this problem, not a claim about which channel matters.
• Intrinsic oscillators and neurovascular coupling. Heart ~1 Hz, breathing ~0.2 Hz, GI tract ~0.05 Hz; these rhythms shape brain dynamics rather than merely being sensed, and hemodynamics may reciprocally modulate cortical gain. Currently modelled as nothing at all, though they “could be integrated as gating signals.”
• Consciousness and metacognition as separate variables. Regulation is largely unconscious, but metacognitive beliefs about one’s capacity to regulate influence both cognition and symptom development — “modeling an individual’s response to internal states and their appraisal may be as important as modeling the signals themselves.” This is Stephan’s allostatic self-efficacy generalized into a modelling requirement; see computational-psychiatry and interoceptive-taxonomy (the accuracy/sensibility/awareness split is exactly this distinction, arrived at from psychometrics).

The translational argument, and its honesty

The clinical section restates the computational-psychiatry programme with one addition the wiki did not have first-hand: computational psychosomatics as Petzschner’s own term (Petzschner et al. 2017), and the computational biomarker as its deliverable — parameter estimates (precision, learning rate, bias) that function “akin to a blood test” for an individual’s maladaptive computation.

What makes the section creditable is the concession attached: “at present there is little empirical work testing such models’ predictions,” and to become clinically relevant the models “will need to be rigorously translated through a development pipeline not unlike that used during novel drug identification.” Effective biomarkers “must demonstrate their utility in improving diagnosis, monitoring, prediction, prognosis, risk susceptibility, or treatment, ideally in individual patients.” That is a higher evidentiary bar than the field usually sets itself, set by the people who would have to clear it — and it matches the roadmap’s own “indirect so far,” three years later and no less indirect.

The comparative-species problem

A section with no counterpart elsewhere in the wiki. There is “a chasm between animal and human studies on interoception,” for two reasons: Craig’s argument that human interoception is qualitatively different from other animals’, and the fact that human interoception research runs on verbal self-report, which no animal can give. Meanwhile the molecular and opto/chemogenetic toolkit that could resolve mechanism is concentrated in one organism, Mus musculus, in which you cannot ask about interoceptive awareness.

The proposed bridge is that computational models yield parameters (precision, learning rate, bias) which are estimable from behaviour in both species without requiring language. The proposed caution is anatomical: parabrachial→insula/vmPFC projections in rats do not exist in monkeys, so more closely related models (macaque, marmoset) are needed. Worth holding against the wiki’s rodent-derived circuit material — survival-circuits, amygdala, stress-induced-analgesia — where the human generalization has been assumed rather than argued.

Berntson & Khalsa (2021), in the same issue, turn this section’s local caution into a general finding with a method behind it. Retrograde rabies tracing shows the cortical networks reaching the adrenal medulla in three synapses in monkey — movement, cognition, affect — are largely absent in rat, whose adrenal network originates mainly in motor cortex. The chasm is not only that we cannot ask a mouse how it feels; it is that the mouse may not have the brain–body connectivity the human theory is about. They also confirm the awareness half flatly: there are no animal models of interoceptive awareness.

The concession, tested within the year

“There is little empirical work testing such models’ predictions” is the sentence this paper is most cited for on the wiki, and it now has a counterexample from the same unit. [[harrison-2021-breathing-interoception-anxiety|Harrison et al. (2021, Neuron)]] — Stephan senior, Petzschner co-author — fitted an associative learning model trial by trial to interoceptive behaviour and regressed 7T activity on its own per-trial prediction and prediction-error estimates.

What it delivers to this paper’s framework, and what it withholds:

• The first arc of the sensory-control-loop gets an empirical handle. State inference from noisy afferents now has model-derived quantities with neural correlates of the right sign — activity falling with prediction certainty, rising with prediction error, in aIns, ACC, MFG and PAG.
• The third arc is untouched. What is learned is a cue-to-resistance contingency, not the body’s own dynamics. forecasting remains the gap page this review named it.
• The Marr discipline holds up, awkwardly. The paper reports that the anterior/posterior dissociation of predictions from prediction errors — an implementational prediction of interoceptive predictive coding — did not appear. That is the level-separation this review insists on, producing exactly the kind of result it warned the field had not looked for.
• The HRL rival is not engaged. Nothing in the design distinguishes an inferential from a valuation account; the models compared were three learning models within one family.
• No computational biomarker. The fitted learning rate predicted neither group membership nor any questionnaire, which is this review’s stated bar going unmet in the first study to try.

Placement

Read as the map of the computational quadrant, standing to the theory what Khalsa et al. (2018) stands to the field. It contradicts nothing in the wiki. What it does is reduce the wiki’s confidence in a framing rather than a finding: predictive coding is an implementation hypothesis with a live rival that reaches the same computational-level answer, and the arc of the loop concerning how bodily states unfold in time is barely modelled by anyone.