the point is that we can be wrong about our emotions not just in the sense that they're not suited to our modern world (that idea has been around for years) but that they're not 'suited' to any world, they don't come pre-packaged and suited to some set of circumstances predicted by evolution — Isaac
Isaac DESTROYS evolutionary psychology. (Maybe).
How I'm thinking about emotions in the natural kind flavour are that they are attractors in the dynamical system of active inference given the statistical regularities of our current lifestyles. So, a dynamical system is a pair of collections, a collection of parameters; called states, like the state of a neuron; and a collection of update rules that maps states to other states; an application of an update rule moves a state "forward in time".
Like if you had the parameter x, and the update rule f(x)=x^2, if the initial value of x is 2, then the updates are f(2)=2^2=4, f(4)=4^2=16, f(16)=16^2=196 and so on. The "time" there is how many times the update rule f is applied.
An attractor in a dynamical system is a collection of states that map into themselves under the update rule. You can't escape it, like a ball rolling to the bottom of a hill. For the above map f(x)=x^2, 1 would be an attractor, as would 0, since f(1)=1^2=1 and f(0)=0^2=0.
A more complicated attractor might be whether an asteroid would enter into orbit around Earth. It'll come from some angle, and when it non-negligibly gets pulled by Earth's gravity, it might start to rotate around Earth. The attractor there would be the collection of all orbits around Earth that the asteroids take.
The collection of states in the active inference model is the collection of states it references, the update rules are the state transitions (what we predict them to be and what our interventions reveal about them intermingling into learning). I'm unclear whether "state" refers to something like the state of a neuron, or whether it refers to something like the state of an environmental parameter, or whether at one stage in the process it refers to an environmental parameter (well, in its encoded form) and at others it refers to neuron states. The active inference system's dynamics also don't seem to have exact state access, like the above square map "knows" that 1 comes in as input, what goes into the update rule in the active inference system looks to be an uncertain summary of each state (from a previous prediction). Anyway.
The system described regarding habit formation in the Friston
paper you linked doesn't have this "gets stuck there forever" property regarding habits though, a prior becomes change
resistant by having its updates diminished by previous success using the policies (actions/worldly interventions, in the paper foraging strategies in a maze) it proposes. So thinking of emotions (not core affect alone) as learned, they would need to be change resistant habits that activate based upon context similarity to the predictions (bodily-environmental model) their representations/encoded patterns generate. When evidence accumulates that the activating context for the habit is no longer present, the agent switches to an exploratory mode that yields the formation of new habits.
If we take that idea that new context recognition is impeded by having a strong prior for what context we're in and what to do in it, it seems to me to fit quite neatly with Barrett's "language-as-a-context" view (from
here, the language paper you linked).
In addition, emotion words cause a perceptual shift in the way that faces are seen. Morphed faces depicting an equal blend of happiness and anger are encoded as angrier when those faces are paired with the word ‘angry’, and they are encoded as even angrier when participants are asked to explain why those faces are angry [19].
Language seems to have the ability to prime which habits are simulated and enacted; and language as a cultural artifact/shared repository of symbols and meanings changes much more slowly than the fleeting associations that shape our emerging experience of emotions. It's a relatively time stable network of associations we partake in by analogous simulations. Moreover, language plays a mediating role in valuation of core affect. So: it changes slowly, it primes for which habits to activate by being a context, it mediates valuation in accordance with its own system of associations. It also seems to amplify predictions/interventions that are more typical of it when it's used as a prime (people primed with angry words report faces as more angry).
That seems to give language the power to canalise the developmental landscape of our emotions. It pulls core affect, through valuation, towards that which it typifies. That makes emotions like "anger", "sadness", despite having variable content, look a lot like attractors to me.