nucleation

LIT-1468: Deeply Felt Affect: The Emergence of Valence in Deep Active Inference

2 min read

Properties5
StatusActive — read 2026-10-23 (NOTE-1111): in full, from the publisher's CC BY PDF as deposited in Aarhus University's repository (pure.au.dk), checked against the University of Amsterdam deposit. A deep active-inference model of valence. A precision term γ sets how far the agent's preference-driven action model, rather than a habitual prior, governs policy selection. Its update, "affective charge", is positive when outcomes raise confidence in the action model and negative when they lower it. A second, slower level infers a binary valence state from affective charge and sends priors on γ back down. Simulated in a T-maze with reversal, the agent with the valence level adapts faster than one without it. "Felt" means explicitly represented, which the authors take as necessary for self-reportable experience of valence. Filed for the record's candidate that a valence regime is felt when the system models its own potential (CLAIM-440).
Tagsemotion-and-affect, probabilistic-modeling, consciousness, behavioral-integration
SupportsCLAIM-438 — An ordering serving as the negotiation's potential, as the potential account of valence defines it, tracks 'wanting', not 'liking': sensitized incentive salience wins a system's conflicts, persists and resumes while hedonic impact measured by taste reactivity stays flat or falls, so the regime the account calls hedonic valence is motivational valence, and hedonic valence, if it is anywhere in the account, is elsewhere, CLAIM-439 — A regime of valence is felt, in the sense of attended and available to the system's other processes, when the system that is in it also models it: when a deeper process infers which ordering wins its negotiation, which way the system is moving along it and how strongly it prevails, and the inference is used elsewhere in the system; unfelt valence is the potential without that model, as unfelt fluency is a change in precision without the inference about it
Cited as “grounds” byCLAIM-468 — Valence intensity need not be the strength of the prevailing ordering's potential: in Hesp et al.'s model negative valence is a weakening of the preference ordering's precision with a fall back to habit, and positive affect is reported to lower a goal's priority, so for either sign stronger valence can go with a weaker hold of the ordering
Cited as “paper” byNOTE-1111 — Hesp et al. — Deeply felt affect: the emergence of valence in deep active inference

Graph View

Casper Hesp, Ryan Smith, Thomas Parr, Micah Allen, Karl J. Friston & Maxwell J. D. Ramstead (2021), Neural Computation 33(2):398–446 — DOI-10.1162/neco_a_01341

Key takeaways

  • Valence is inferred from confidence in one’s action model. “agents infer their valence state based on the expected precision of their action model—an internal estimate of overall model fitness (“subjective fitness”)” (p. 399).
  • Affective charge (AC) is the update to that precision. It “lends a sign to otherwise unsigned divergences between predictions and outcomes” (p. 399). It is zero when policies go as expected and positive or negative as outcomes raise or lower confidence (pp. 409–410).
  • What γ does. “higher γ values afford a greater influence of the expected free energies of each policy entailed by one’s current action model” (p. 408), against a fixed habitual prior over policies. With low γ the agent will “fall back on her baseline prior over policies” (p. 420).
  • The deep level. A second level, slower and across trials, holds a binary valence state. It takes AC as evidence and sets the lower level’s prior on γ (pp. 416–420). The authors stress that they “do not equate this notion of expected precision (or confidence) with valence directly” (p. 426).
  • “Felt” means explicitly represented. “Presumably this type of explicit valence representation is also a necessary condition for self-reportable experience of valence” (p. 429). And “it is an open question which level or subset of levels may be privileged with respect to its contribution to affective phenomenology” (p. 433).
  • A simulation, not data. A T-maze over 64 trials with reversal at trial 33. The valence level speeds re-adaptation and attenuates repeated AC. Before reversal, “their behaviors were indistinguishable” (p. 425).
  • Its own hardest problem. It names “the strongest and most important challenge” (p. 430): dopamine, a candidate carrier of AC, is tied to “wanting” rather than pleasure.

Standing in the record

Filed on 2026-10-23 for the record’s candidate that a valence regime is felt when the system models its own potential (CLAIM-440), after Brouillet and Friston (NOTE-785, THEORY-662), who cite this paper for valence. Structurally, it is the nearest existing model of that candidate. γ is the weight of one ordering against another in policy selection, and the deep level models it. But it identifies valence with the self-model, not with the potential. And it makes negative valence a weakening of the preference ordering’s potential. Read in NOTE-1111.

Backlinks