[1]
The following quotations might help to elucidate the matter: “[f]eelings typically express affect and valence in sensation (25-26), all the feelings vary in affect in roughly the same way, because they all include valence in their informational structure” (p. 20).
[2]
In Proust’s words, the difference between “hot” feelings and feelings with valence, on the example of M-feelings, is that “although all M-feelings do not often have a definite ‘hot’ quality comparable with fear and love, they always have a valence, according to whether they predict the agent’s progress toward or away from her cognitive goal” (p. 21).
[3]
To be more precise, the question is about the functional description of the formal object of feelings. Proust (this collection) says that “[f]eelings express […] affordance as their focus (for formal object), along with its graded valence, ranging from very unpleasant to very pleasant, and with its intensity gradient, which ranges from small to large” (p. 8). Affordance is defined as “perceived utility”, and can be positive or negative (ibid., p. 7). Positivity and negativity are dimensions along which valence changes, and valence has been characterised as the rate of change of discrepancy towards the (cognitive) goal. For more on why the latter characterisation is interesting see section 3.
[4]
Barrett & Bar (2009) define affect as an influence on bodily states that is either unconscious or, if conscious, experienced as pleasurable or unpleasurable to varying degrees (pp. 1327–1328). Barrett & Bar’s (2009) basic claim is that the orbitofrontal cortex (OFC) integrates into a unified multimodal representation sensory information from both world and body in a dynamic way.
[5]
One could also ask whether the same evidentiary boundaries would be involved in feeling and perceiving, since there could be many of them (Friston 2013).
[6]
Among those who agree with Proust that the content of epistemic feelings is non-conceptual and non-metarepresentational are, for example, Michaelian & Arango-Muñoz (2014). But the content of a metacognitive feeling being non-conceptual does not preclude that concepts play a causal role in its emergence.
[7]
A better understanding of the indexical mode of feelings might be provided by the following quotation: “Feelings can be seen as pre-specified states of a comparator, which predict ultimate success or failure in the actions that they monitor. Given that the information they carry is immediately used in controlling and monitoring current effort, it is misleading to present them as ‘reporting’ the epistemic properties of a mental state or referring to it (even de re). They are, rather, signals in a control mechanism, which work somewhat as traffic lights do: allowing traffic, stopping it, rechanneling it; no report or reference need be involved” (Proust 2013, p. 76). In another place Proust (2013) notes that feelings “do not properly ‘refer’, because they do not engage propositional thinking” (p. 77).
[8]
Note the analogy to the “dark room problem” in predictive coding: if an agent wants to minimize surprise or prediction error, then she should stay in a dark room, given that there will be no surprise in it (e.g., Clark 2013). If there were no prediction error, this would cause uncertainty (e.g., Friston et al. 2012). Proust’s argument is similar: if there were no violations of expectations, then metacognitive feelings would not have any valence, because they only have valence if the rate of change is quicker or slower than expected.
[9]
Attention is precision optimization according to predictive coding (Hohwy 2013).
[10]
Mathys et al. (2011) are also interesting for the given topic insofar as Proust argues that the heuristics upon which metacognitive feelings are based might be changed via associative learning; Mathys et al. (2011) provide a predictive coding model of reinforcement learning.
[11]
For a predictive-coding model of a goal-directed action see Friston et al. (2013).
[12]
Emotional valence has been also argued to be modelled as the rate of change of free energy: Instead of estimating volatility or “slow and continuous changes in states of the world” the rate of change of free energy is argued to take that role of estimating (known) uncertainty (Joffily & Coricelli 2013, p. 1). Here Joffily & Coricelli (2013) accept Yu & Dayan’s (2005) distinction between expected and unexpected uncertainty: Expected uncertainty is the one about known unreliability of predicting relationships within a context and unexpected uncertainty is the one about the appropriateness of the context itself such that when unexpected uncertainty is high, it is a signal that a context switch should be made.
[13]
For an elaboration on the phenomenal signature of knowing in intuitions of certainty, see Metzinger & Windt (2014).