Predictive processing: Unification by fiat
Predictive processing (PP) theory of cognition claims that nervous systems continuously try to anticipate their sensory inputs, following the imperative to minimize the mismatch between their predictions and sensory activations. In the process, the internal model of causal dependencies underlying sensory signals is developed and refined. The theory has recently sparked great hopes that it will provide a unifying account of the entirety of human and non-human cognition: In this view, virtually all cognitive processes are guided by the predictive nature of the internal model. In our recent paper (written by myself and Marcin Miłkowski) we argue that this enthusiasm is premature: PP faces serious problems which effectively stunt its development as a scientific theory. The paper was recently published in the journal Cognitive Science and is now available in open access. It is our second criticism of overextending PP (the previous one is a commentary in BBS on Gilead et al. model of abstract thought).
Our skepticism is based on two main observations. In the first part of the paper, we focus on the “theory-implementation gap”, i.e., the gap between an elaborate mathematical apparatus and its interpretation in terms of psychological phenomena or neurobiologically plausible cognitive architectures. While the formalization of PP is quite strict (but also, notably, very diverse, as there are many different PP algorithms), the theoretical interpretation of its core terms seems to be free-for-all. For example, precision—a concept related to the variability of the signal; thus, to the computational “confidence” or “trust” in its reliability—is being identified with many distinct cognitive and psychological phenomena, including subjective feelings of confidence or trust. This observation applies also to other core technical PP terms, such as predictions or beliefs. Explanations based on such identifications are non-informative, as they are based on the fallacy of equivocation—the confusion of several meanings of a single term in an argument. The theory-implementation gap surfaces also in the computational models of psychiatric phenomena, which do not abide by the root commitments of the theory, regarding functioning and structure of the predictive model.
In the second part of the paper, we present arguments that PP models are not validated empirically in a sufficiently rigorous manner. The vast majority of PP modeling work takes the form of theoretical re-descriptions of contemporary models. These accounts neither yield new, testable predictions nor enable new inferences on phenomena being ‘covered’; they simply use generic PP language to rephrase existing models in hypothetical predictive terms. Moreover, PP proponents harness certain argumentative strategies to conceal the limited usefulness of their models which are often defined as “starting points”, to be empirically validated in the ‘foreseeable’ future. However, they almost never are; instead of testing them against viable alternatives or outcomes explicitly incompatible with the theory, proponents of PP succumb to consistency fallacy, taking selected (broadly consistent) evidence as ‘providing support’ for PP. Not only such practices do not corroborate the theory, but, as we show in the article, they also lead to mutually exclusive PP models of the same phenomenon and wishful interpretations of the obtained data (even in cases when they seem to undermine PP).
The upshot is that, as for today, PP is definitely not a great unifying theory it was promised to be and its development is arrested. In particular, it displays two features that unifying theories should not possess: It is heterogenous and unsystematically applied. Numerous diverse interpretations of technical terms result in mutually exclusive accounts of the same phenomena or theoretical and computational models being at odds with theory’s fundamentals. Therefore, PP is either underspecified or contradictory itself, but it is certainly not unifying.