NOTE-434: The theory of constructed emotion: an active inference account of interoception and categorization

Contribution

The paper restates Barrett’s conceptual act theory, renamed the theory of constructed emotion, from the brain up rather than from emotion words down. It derives the theory from three premises: the brain’s job is allostasis, it does that by prediction, and its predictions are concepts. It then maps the architecture onto cortical cytoarchitecture (Barbas’s structural model) and onto three intrinsic networks, and lists the neuroscience it takes to disconfirm the classical view (Table 1) and to support its own (Table 2).

Key insight

Perception, action and emotion are one process. A brain that must regulate its body before needs arise has to predict, and to predict it has to use past experience to categorize what is coming. “Predictions are concepts. Completed predictions are categorizations”. When the concept doing the categorizing is an emotion concept, the result is an instance of that emotion. So an emotion is “constructed the same way that all other perceptions are constructed”, and “Emotions are constructions of the world, not reactions to it” (Conclusions).

Assumptions

• The brain’s core task is allostasis (“What is a brain for?”), following Sterling: anticipating and meeting the body’s needs. Regulation of autonomic, immune and endocrine systems runs under every other function.
• Predictive coding as the brain’s computational scheme (“An internal model is predictive, not reactive”). The paper calls it “an increasingly popular hypothesis” and treats it as the frame.
• Barbas’s structural model of corticocortical connections. Prediction signals flow from the deep layers of less laminated (agranular, limbic) cortex to the upper layers of more laminated cortex. Prediction error flows the other way and is compressed on the way.
• Degeneracy and population thinking (Edelman & Gally, after Darwin’s Origin as read by Mayr). A category is a population of variable instances grouped by a goal, and its summary representation “need not exist in nature”.
• Perceiver dependence. Emotion categories are conceptual categories of “social reality”, real in the way money is (point 1 of the list of claims not made; footnote 15).

Key results

The paper reports no new experiment. Figure 3 (simulation BOLD, N = 16, from Wilson-Mendenhall et al. 2013) and Figure 6 (a resting-state allostasis and interoception system, from 280 participants and replicated in 270, from Kleckner et al.) are the closest to original data, and both are cited from other papers. What it gives is the theory’s structure:

• Affect and interoception. “Interoceptive sensations are usually experienced as lower dimensional feelings of affect”, and valence and arousal are “basic features of consciousness … that, importantly, are not unique to instances of emotion”.
• Concepts as whole-brain predictions. A concept is “a collection of embodied, whole brain representations that predicts what is about to happen in the sensory environment, what the best action is to deal with these impending events, and their consequences for allostasis” (Glossary). Concepts are ad hoc and goal-based. “happiness” is constructed anew each time relative to the situation’s goal.
• Architecture (Figure 5):
  • Agranular limbic regions (ACC, ventral anterior insula) send visceromotor predictions to the hypothalamus and brainstem, with efference copies to motor cortex and all sensory cortices.
  • Primary interoceptive cortex (mid/posterior insula) sends sensory predictions to the better-laminated exteroceptive cortices, and motor cortex sends its own (hypotheses, so labelled).
  • The amygdala signals uncertainty, not emotion, and ventral striatal prediction errors mark allostatically relevant input.
• Networks. The default mode network is “necessary for the brain’s internal model”, hosting multimodal summaries. The salience network sets precision, choosing which prediction errors to learn from. The frontoparietal network holds simulations over longer timescales. Hence “Emotion perception is event perception, not object perception.”
• Against the classical view (Table 1). Emotion categories cannot be consistently localized to regions or intrinsic networks. MVPA patterns for a category need not occur in any instance. Intracranial stimulation finds no emotion-specific receptive fields. Amygdala lesions have variable effects (patients SM, AM and BG, Box 1). Rodent “fear” circuits show many-to-one and one-to-many mappings.
• Pattern classifiers misread. Classifiers that separate emotion categories do not replicate across studies, and a classifier pattern is “a statistical summary of a highly variable set of instances”, not the brain state for the emotion. Taking it as such is an “essentialist error”.
• Mental inference fallacy. Equating an action, such as freezing, with an emotion such as fear. The paper says this “has wreaked havoc” with the accumulation of knowledge. Decorticate “sham rage” is pattern generators without the conceptual system.
• Appraisal theories. Because “meaning does not trigger action, but results from it”, classical appraisal theories that “assume that a response derives from a stimulus that is evaluated for its meaning (e.g. Lazarus, 1991; Scherer, 2009; Roseman, 2011)” are “highly doubtful”. “Appraisals as descriptions of the world (e.g. Clore and Ortony, 2008), however, are produced by categorization with concepts”.
• Non-human animals. “‘Is the fly fearful?’ is not a scientific question”, but whether a human perceives fear in a fly (“yes”) and whether the fly feels fear (“no”) are. The fly “may feel affect”.

Claims

id	claim	strength	support
C1	The brain’s core function is allostasis, achieved by an internal predictive model whose interoceptive part is felt as affect	moderate as a synthesis; cited (Sterling; Barrett & Simmons 2015; Seth 2013)	“What is a brain for?”, “How does a brain perform allostasis?”
C2	Predictions are concepts, and an instance of emotion is a categorization with an emotion concept	theoretical; the paper’s central hypothesis	”The theory of constructed emotion”
C3	Prediction signals originate in agranular limbic cortex and flow down a laminar gradient; prediction errors flow up and are compressed	moderate: anatomy cited (Barbas); the computational reading is hypothesis	Figures 4A, 5; footnote 16
C4	Emotion categories have no consistent neural signature, by region, network, pattern or neuron	moderate: cited meta-analyses and studies, Table 1	Table 1; Clark-Polner et al. 2016
C5	Successful emotion classifiers do not replicate across studies and do not identify emotion brain states	moderate: cited, plus an argument from Posner & Keele	”Selected implications”
C6	Causal appraisal theories are doubtful because meaning results from action and does not trigger it	argued from C2–C3, not tested	”Computational architecture …”
C7	The amygdala signals uncertainty and adjusts allostasis, and is not an emotion centre	hypothesis	”Computational architecture …”, Box 1
C8	Reappraisal and other regulation are accomplished with predictions that categorize sensory input	hypothesis	Conclusions; Figure 4C
C9	The theory gives a common frame for mental, physical and neurodegenerative disorders, and for chronic pain	asserted, cited to other Barrett papers	Conclusions

Concepts

• allostasis: “Regulating the internal milieu by anticipating physiological needs and preparing to meet them before they arise”.
• interoception: the representation and use of sensations from the internal milieu; “at the core of the brain’s internal model”.
• affect: valence and arousal, the felt consequence of allostatic predictions about the body. Present in every waking moment, not only in emotion.
• concept / category: a category is a population of events treated as similar for a goal in a context, and a concept is the population of representations of them (footnote 15). Here a concept is a cascade of whole-brain predictions.
• degeneracy: “the capacity for biologically dissimilar systems or processes to give rise to identical functions”. Distinguished from redundancy.
• classical view of emotion (Figure 1): basic emotion theories, causal appraisal theories, and black-box functionalism. Each assumes that every emotion has an essence: a Lockean one (an underlying mechanism, neural or appraisal-based) or a Platonic one (a fingerprint in face, physiology or appraisals).
• mental inference fallacy: treating an observed action as evidence of a specific emotion.
• precision signals: predictions of which prediction errors are worth learning from. The paper assigns them to the salience network.

Connections

• Lazarus, Emotion and Adaptation (LIT-550), Deferred. Named as the first example of a causal appraisal theory (claim C6). The paper engages the type, not the book’s specific arguments, and quotes nothing from it.
• Frijda (LIT-540, read; LIT-546, Deferred). Not cited. Frijda’s late statement agrees with this paper on several points. Motives are “set to expect a particular input”, as in predictive coding; concern pertinence is appraised automatically; and actions are anticipated through their exteroceptive and interoceptive sensory consequences. But it keeps appraisal as what elicits action readiness, and keeps modes of action readiness that “correspond to certain common emotion names”. On this paper’s terms, that is a Lockean essence.
• Gross (LIT-554), read. This paper cites Gross (2015) and Gross & Barrett (2011), not the 1998 review. Its hypothesis C8 keeps reappraisal and gives it a mechanism: changing the predictions that categorize. It does not keep the 1998 model’s sequence, in which situations generate meanings and meanings generate responses.
• Seth (LIT-135; NOTE-169), read. Both papers take interoceptive predictive processing that serves allostasis as the base layer, with shared sources (Seth et al. 2012; Seth 2013; Seth & Friston 2016). Seth’s §4.1 grounds consciousness in it, while this paper grounds affect in it and calls affect a basic feature of consciousness. The two are compatible. Neither paper tests the other’s further step.
• Friston (LIT-526; NOTE-421), read. The paper’s active inference is the neural predictive-coding scheme (Friston 2010; Adams, Shipp & Friston 2013; Shipp et al. 2013 on agranular motor cortex), not LIT-526’s Markov-blanket theorem. Nothing in this paper depends on LIT-526’s claim that any blanketed ergodic system “appears” to infer. Footnote 5 grounds the need for an internal model in Conant and Ross Ashby’s good-regulator theorem instead.
• Damasio and Carvalho (LIT-388; NOTE-330), read. Cited for the interoceptive pathway, which ends in “granular layer IV of the primary interoceptive insular cortex”. NOTE-330 records Damasio’s action programmes (fear, anger, sadness and the rest) as innate and distinct from feelings. That is the categorical, essence-bearing picture this paper rejects, while it shares Damasio’s view of feeling as the read-out of body state. Damasio (1999) is also among the authorities cited for affect as a basic feature of consciousness.
• Somatic markers (LIT-384, LIT-383). Box 1 reports that patient SM shows no anticipatory skin-conductance response in the Iowa Gambling Task, the measure LIT-383 tracks in ventromedial patients. The paper’s reading is that her brain “seems not to use arousal as a learning cue”, not that it lacks somatic markers. The lesion is the amygdala, not the ventromedial cortex of LIT-384, though the paper notes SM’s damage extends to vmPFC.

Bearing on the record

• It sources one side of a THEORY proposed in this batch’s report: that emotion categories are not natural kinds with neural essences, but populations constructed by categorization. That stays Proposed while the record holds only the theory’s own statement and no independent test or reply. Adolphs’s companion piece (nsw153) is the obvious reply to file.
• It bears on the interoception cluster. LIT-135, LIT-388 and this paper agree that interoceptive, allostatic prediction is the layer feelings report. That agreement could ground a THEORY of affect as felt allostatic prediction, sourced from all three, which is also proposed in the report.
• No instruction for machine-learning practice.
• It is the source of THEORY-057, which rivals THEORY-055. With LIT-388 it sources THEORY-045, and with LIT-540 THEORY-056.

Limitations

• Hypotheses are the body of the paper. “I hypothesize” introduces most architectural claims: the motor-to-sensory and insula-to-sensory predictions, the amygdala as uncertainty signal, the roles of the networks. The paper calls itself “a mere sketch”.
• The case against the classical view rests on disconfirming essentialist readings. It is weaker against appraisal theories that never claimed a neural fingerprint per emotion. The paper’s Figure 1 nonetheless counts causal appraisal theories as positing a Lockean essence (“a set of evaluative mechanisms called ‘appraisals’”), and does not engage any specific appraisal theory’s predictions.
• The positive evidence (Table 2) is largely consistency. Degeneracy and neural reuse are predicted by the theory, but they are also compatible with many other non-modular accounts.
• Predictive coding is assumed, not argued. The theory inherits that framework’s open questions.

Open questions

• What would disconfirm the theory? The paper says new paradigms raise “more questions than they answer”, and lists no test that the classical view passes and constructionism fails.
• Can a constructed emotion concept be shown to change interoceptive and physiological prediction, and not only report? Figure 3’s simulation result is suggestive.
• How does the account handle appraisal theory’s best-supported finding, that appraisal profiles predict which emotion people report? The paper’s answer, that appraisals are descriptions produced by categorization, is stated, not tested.

Corrections

• Author’s correction (nsx060, 2017). The degeneracy sentence originally read “in different contexts”, and it should read “in the same context (i.e. many-to-one mappings of structure to function)”. The PMC text read here has the corrected wording.
• Pagination. Crossref gives no page range (article id nsw154). PubMed gives 12(1):1–23. The correction notice cites the article as “12 (1): 17–23”, which conflicts with PubMed. This record uses 1–23.