Self, craving and active inference
I recently read a paper that seeks to explain, within the framework of active inference (see my posts on active inference here and here), the self, (Buddhist) craving, the connection between them, and the effect of meditation on them.
I'm not sure I understood everything, but here's my interpretation of the paper:
We have a meta-system that tracks the evolution of prediction errors, and our capacities and performance in reducing them. This meta-system allows to judiciously adjust the distribution of precision (importance) to policies (action plans serving goals).
This meta-system infers the existence of a “self” which is the possessor of goals, the possessor of capabilities, and which serves as the “hidden cause” of actions and the central entity in counterfactual reasoning for evaluating policies.
This explains several aspects of the self. For example, the fact that the self is conceived as the hidden cause of actions and as the possessor of capabilities explains the “feeling of agentivity” associated with the self.
In this model, affect (hope, satisfaction vs. frustration, disappointment) depends on whether the meta-system judges the self to have good control/good performance (i.e. is good at self-evidencing): poor control by the self = negative affect; good control by the self = positive affect.
When affect is negative, the meta-system intervenes on the precision of policies to try to correct this.
As I said earlier, the meta-system assigns goals to the self, but also policies. For example, “I'm an agent who must become a baseball champion” (goal), but also “I'm an agent who scratches when it itches” (policy).
Indeed, the meta-system, by often observing the organism scratch when it itches, as it sees the self as a hidden cause of actions, infers that this policy “is part” of the self.
The self is also assigned the goal/policy of avoiding pain: “I am an agent who avoids/should avoid pain”.
Now, what happens when we deviate from a goal/policy that's part of the self? The meta-system panics:
That threatens the integrity of the self model (“I'm an agent who scratches when it itches, but it itches and I don't scratch; error!”).
The meta-system thinks that the self is not in control (“what's the self doing, it's supposed to scratch when it itches, it has no control!”).
The meta-system sees that prediction errors are not being addressed (“it itches and it keeps on itching, we've got to do something!”).
The meta-system “has to protect” the self (the self being a kind of center of gravity or representation of all this survival machinery), and since it models the self as having this goal/policy as its essential part, it perceives this deviation not just as a deviation but as a “threat”.
The meta-system starts to "want" to address this deviation, it forms expectations (predictions) about the resolution of the deviation, and if the expectations are not fulfilled...
Etc.
In a sense, one of the main functions of the meta-system is the “contextuality” it enables; we shouldn't allocate precision to our itchy nose if a bear is chasing us. But this contextuality =/= “rationality”. The meta-system postulates and constructs an entity (the "self") which it judges and protects, and if it judges that the self is doing badly, or that something is a threat to the self, the meta-system goes into a panic (that's its function). His judgment is contextual, but his judgment and reaction are irrationally dramatic and exaggerated.
What happens, then, is a negative affect, and on top of that, the meta-system cranks up the precision of certain policies to address the concern, even though the concern may be minimal (a slight itch).
Phenomenologically, it's the Buddhist “craving”; a strong, negatively-charged emotional impulse with an almost existential aspect (“things shouldn't be like this, I'm dying! crying out in anger”).
Pain is the simple deviation from a goal. Craving is something that begins to exist when you have a meta-system and a self. This deviation takes on a negative affect, this deviation is seen as a threat to the self, to its control, etc., and there is a huge increase in the precision of policies to address this deviation.
Buddhism says that we can have pain without it making us suffer, that we can have unfulfilled desires without it making us suffer, that we can lose attachment to pleasure and fear to pain...
How do we reach this state? By meditating (see my post on meditation).
By meditating, the meta-system realizes that sensations are impermanent (they emerge and then dissipate without the need for intervention); it realizes that they are not essential to the self. While meditating, we practice inaction (we refuse to act according to the policies selected by the system), and the meta-system realizes that policies are not essential to the self. By meditating, we examine the various aggregates of consciousness, distancing ourselves from them (making them “opaque”), and the meta-system realizes that they are not essential to the self. By meditating, we learn to control the assignment of precision. Etc.
By meditating, our meta-system stops modeling the self as a rigid entity that must be protected, stops seeing pain or goal deviation as an existential threat to the self. It stops panicking at the slightest deviation. The meta-system stops attributing all kinds of identities to the self: “I'm an agent who is x”, “I'm an agent who likes x”, “I'm an agent who does x”… The meditator can himself control his precision assignments (this control is pleasant, the meta-system likes the feeling of control). Etc.
The meditator continues to feel pain and pleasure, but loses these strong, emotionally-charged urges. Pain and pleasure are no longer “great danger for him”, they are just useful, emotionally neutral information. And at the same time, the meditator reports an experience of no-self, in that the meta-system has stopped seeing things as essential to the self.