predictive processing

The theory of predictive processing seems to be attracting a lot of interest in neuroscience and meditation circles. I want to try to understand this. In particular, some questions:

• What exactly does the theory say?
• What testable predictions does it make?
• How does it relate to the standard Bayesian decision theory account of rationality?

The theory says that the brain is constantly making predictions about sensory data, and that the goal is to minimize prediction error. Broadly there are two ways to minimize prediction error:

• Update your model of the world to fit the sensory data you're seeing.
• Act to change your sensory data so that it matches the prediction. This could be by simple manipulation of your sense organs, e.g., opening your eyes, but includes more broadly any action that you take to affect the state of the world.

Why would you prefer one of these routes over the other? Some aspects of the world are easier to change than others. Conversely, the theory as I understand it also seems to require that predictions have varying levels of 'stickiness' - some predictions, like the prediction that we have a body, are deeply baked in and harder to change than others. A prediction that is too sticky to change can only be satisfied by taking action in the world, so it functions essentially as a goal (e.g., to maintain the integrity of the body).

References that seem useful:

• https://marksprevak.com/publications/predictive-coding-i-introduction-40b2/

Technical questions:

• What does 'prediction error' ground out to in terms of concrete loss functions? Is this necessarily a probabilistic loss, i.e., weighted by confidence in the prediction?