Testable or bust: theoretical lessons for predictive processing

In our new paper – just out in Synthese – we tackle a long-standing question whether predictive processing (PP) is actually testable. This is our take.

We first clarify the conceptual territory by distinguishing many scientific representations (toolboxes, frameworks, models, and theories) that may be branded as PP. Toolboxes and frameworks cannot be tested, and, as such, should be put aside when discussing falsifiability. (2/n)

Thus, we focus on PP as a theory which may be understood either in its ‘generalized’ or ‘hierarchical’ version. Only hierarchical PP entails non-trivial implementational commitments, and may be assessed for theoretical virtues, such as testability.

In the context of testability/falsifiability, theories face two kinds of problems: counterexamples (Popper) and incorrect models (Taatgen), of which the latter are actually much more important, as they show that theories are insufficiently restrictive to disallow false models.

In result, they are capable of generating explanations for any conceivable phenomenon as envisioned by current scientific consensus. When the consensus changes, such theories change as well. At some point, they may even “predict” impossible or already debunked phenomena.

This is currently the case with PP, as we discuss in the context of ‘what’ and ‘where’ visual streams, self-deception, and complete class theorem which may be largely responsible for the modeling mess.

We conclude that PP is currently used as a computational framework. It is absolutely fine, but frameworks should be useful and productive, and not in the constant need of ad hoc tweaks and patches. They can also still be assessed for constraints, such as computational tractability.

If PP wants to grow as a theory, the field has to focus on the implementation details of hierarchical message passing theory, as this is what makes PP experimentally contentful and constrains possible PP-based explanations, both on behavioral and neurobiological levels.