First, I would like to repeat, in short, the main claims of the target paper that will serve as a basis for my subsequent comments and extensions in the following sections. Joëlle Proust’s article is concerned with the functional and informational characterisation of feelings. She argues that the concept of “feeling” consists of the following components:
Reactive (associated with appraisal)
Subjective experience
With distinctive embodied phenomenal quality (somatic markers have the function of expressing intensity and valence of feelings, Proust this collection, p. 8)
Possessing a formal object (not always, e.g., feeling depressed; absence of a formal object is typical of moods, footnote 5)
The formal object of feelings is argued to be affordance-sensing, a “non-conceptual and entirely subjective appraisal of the environment by the agent” (Proust this collection, p. 7) or a “subjective, embodied relation to an opportunity in an input from the environment” (p. 8). Assuming the non-referential indexicality of feelings, or that feelings signal a relational affordance (p. 7) that depends on the representational system (p. 8), Proust argues that feelings can misrepresent only if they are re-described in propositional terms. She argues that feelings are transparent, because of the experienced connection between their somatic markers and affordances, as well as because of the direct influence of their valence and intensity on an agent’s motivation (p. 12). Though subjects feel directly, in order to report their feelings they have to “translate one mode of representation into another, with no guarantee that this translation will not enrich or modify FS intentional content” (p. 12). Subjects might reinterpret and mis-describe their feelings, but they cannot thereby change the nature of those feelings (feelings being cognitively impenetrable; p. 19).
Feelings are argued to be a plausible candidate for a natural kind on the basis of the comparison between feelings and emotions—which she considers not to constitute a natural kind (Proust this collection, p. 3). Two kinds of subjective appraisal might be part of an emotion: primary feelings on the one hand and appraisals cum conative dispositions on the other. While the first kind corresponds to an earlier time in our evolutionary development, is independent of concepts, induces specific responses, and possesses distinct somatic markers, the second kind is not and might be a blend of different instances of the first kind. Apart from primary affective feelings, somatic, agentive, and metacognitive feelings are argued to form a natural kind.
The function of feelings is to non-conceptually evaluate and signal the result of a comparison process between prediction and outcome through embodied experience (Proust this collection, p. 4). Due to their non-conceptual monitoring nature, feelings do not convey, but merely approximate a causal relation between internal states, external states, and actions (p. 11). There are three kinds of functional relations between feelings and actions (pp. 4–5):
Determination of a kind of action in response: approach vs. avoidance
Specific orientation in time: predictive vs. retrospective
Level of arousal: elevation in energy vs. soothing effect
Feelings are argued to be the result of a comparator or control mechanism that is automatic and encapsulated. The latter requirements are imposed in order to explain the independence of feelings of beliefs and intentions (p. 5) such that, e.g., one could still feel the adrenalin rush even though the hypothesized venomous snake turned out to be a twig.
Metacognitive feelings (M-feelings) are held to express informational, instead of environmental affordances, arise in mental acts, and trigger similar actions of approach or avoidance. M-feelings involve appraisal of the properties of the informational processes underlying contents of thought, but not those content themselves. Against Schachter & Singer’s (1962) two-factor theory of emotions (interpreted as feelings possessing intrinsic arousal but extrinsic valence), Proust argues that feelings have intrinsic intensity and valence. Cues on which those feelings are based can be conveyed verbally though, and thus, the heuristics (implicitly or explicitly) might change in the long run. The main claim is thus that context-dependency is not concept-dependency (Proust this collection, p. 17). Experience of tokens of feelings differs with respect to the kind of affordance they express (several affordances might be linked to the same type of feeling) and actions they trigger.
An especially interesting claim for me is that affective feeling in general, and metacognitive feelings in particular, have a meta-monitoring function of signalling “the rate of reduction of discrepancy toward a confidence threshold” (Proust this collection, p. 21). If the rate of discrepancy reduction is above expected, the valence of a feeling is experienced as more positive, and, if below expected, as more negative. “Cold” feelings without valence are those for which the expectation has been correct. This claim is interesting for two reasons. On the one hand, to the reader familiar with the self-deception literature the key-concept “confidence threshold” will stand out. It plays an important role in accounts of self-deception that regard it as a kind of hypothesis testing (one prominent proponent of this view is Mele 2012). In short, according to this type of account, gathering of evidence in favour of a certain hypothesis is pursued up to a certain point: up until the amount of evidence has reached a confidence threshold that is enough to push an acceptance or rejection of the hypothesis (for more see Pliushch & Metzinger 2015). On the other hand, “prediction error”, or difference between prediction and sensory input, is the key-term in the model of mental representation that has lately gained a large amount of acceptance—predictive coding (for a short introduction to the free-energy principle of which predictive coding is a particular implementation see Friston 2009; see also Clark, Hohwy, Seth this collection). Predictive coding provides a unifying explanation for perception, cognition, and action as a result of hierarchical Bayesian inference: at different levels, predictions are compared to propagated precision-weighted prediction error that, under different conditions, leads either to changes in the model of causes of sensory input or to action directed at testing the current model (Clark 2013).
The idea that feelings signal the rate of reduction of prediction error might be worth elaborating in the predictive coding framework, particularly given the recent study by Furl et al. (2010) who argue that facial expressions are represented as anticipated trajectories of the change of those expressions: pictures of neutral and fearful faces were morphed to different degrees such that participants got to see trajectories from a neutral to a fearful face and vice versa. After seeing such a sequences of pictures, participants had to rate another picture for fearfulness. The results indicated that predictable sequences in which the degree of being morphed rose or fell monotonously, thus forming a trajectory, biased perception (Furl et al. 2010, p. 696). Combining Proust’s idea with the results of Furl et al.’s study: feelings might also be represented as anticipated trajectories of change, particularly given the possibly bi-directional causal influence between feelings and facial expressions (see section 2_2).