Introduction. Predictive processing (PP) holds that the brain is a hierarchical generative model that continually predicts its sensory input and propagates only prediction errors upward, with precision-weighting deciding which errors are allowed to drive revision. Perception is inference; action is the other way of reducing error, by changing the world so that it matches the prediction. Friston’s free-energy principle (FEP) generalises this: any system that persists in a changing environment must, in the long run, appear to minimise a bound on surprise, variational free energy, and its boundary is formalised as a Markov blanket—a set of states that renders internal and external states conditionally independent. Bruineberg and colleagues traced the blanket from Pearl’s Bayesian networks, where it is a formal device inside a model, to active inference, where it is asserted to be the real boundary of an organism, and showed that the two uses—”Pearl blankets” and “Friston blankets”—have been run together. The first is well supported and does little metaphysical work; the second does the work but needs premises the mathematics cannot supply. A parallel caution applies to the FEP itself: as a variational principle it is close to analytic, and the empirical content lies in the auxiliary claims.
Important authors. Karl Friston (Wellcome Centre for Human Neuroimaging, University College London; FRS) originated the FEP and active inference. Andy Clark (Sussex) and Jakob Hohwy (Monash University) are the philosophers most responsible for predictive processing’s uptake. Jelle Bruineberg (Macquarie University), Krzysztof Dołęga, Joe Dewhurst and Manuel Baltieri wrote the Behavioral and Brain Sciences critique. Rao and Ballard’s 1999 model is the computational origin of cortical predictive coding.
Importance for cybernetics and the VSM. PP descends from Ashby and Conant–Ashby: a regulator that models what it regulates. Active inference reproduces several VSM functions from one principle—prediction-error minimisation across a hierarchy of timescales resembles the Three–Four homeostat, precision-weighting resembles variety attenuation, and “change the world” gives the niche-construction move Beer lacks (T8.5). There is no separate coordination organ in the FEP. If it can generate System Two’s function from the principle that generates Three’s and Four’s, the five functions are a coarse partition of one process rather than a minimal set (Applying new lenses §III). The VSM literature has not engaged PP; the FEP community, by contrast, absorbed the blanket critique in its most visible venue and continued.
Importance for the article. Three places. §2.2 adopts Pearl versus Friston blankets as the organising binary—instrumental versus constitutive readings of the five systems—and states that all three v01 demonstrations tried to test constitutive claims with evidence that licenses only instrumental ones; the rivals in every later receipt are written so the two readings predict different observations (from Input from Cognitive Science §1). §2.3 uses the FEP as the parallel case of a formally true core surrounded by empirical claims that borrow its authority, so that the theorem/operationalisation problem in Demonstration II (§7) is not parochial (Applying new lenses §III). §13.3 cites the FEP community’s absorption of Bruineberg et al. as a precedent for corrigibility without a register. Reviewers will press on whether the instrumental/constitutive binary is exhaustive, on whether a receipt can really keep the two readings apart in practice, and on the unification claim: if PP subsumes S2–S4, what is the discriminating observation? Note that Clark (2013, 2016) are Tier C and not cited in v02’s text; Bruineberg et al. is Tier A.
Sources in the reading list.
- Pearl versus Friston blankets, inference with a model versus within a model; the source of §2.2’s binary, read in full including the commentaries and response.
- the hierarchical prediction machine and its difficulties in one article; read for how precision-weighting and the action route are defined.
- the book-length version; read the chapters on action and on the embodied, situated agent for the “change the world” route.
Other important sources and authors.
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138 — the standard short statement of the FEP from its author.
- Friston, K. (2013). Life as we know it. Journal of the Royal Society Interface, 10(86), 20130475 — where the Markov blanket is made the boundary of a living system; the paper the Friston-blanket reading rests on.
- Hohwy, J. (2013). The Predictive Mind. Oxford University Press — the other philosophical synthesis, more internalist than Clark’s; useful for the contrast.
- Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87 — the computational model that PP generalises.
- Raja, V., Valluri, D., Baggs, E., Chemero, A., & Anderson, M. L. (2021). The Markov blanket trick: On the scope of the free energy principle and active inference. Physics of Life Reviews, 39, 49–72 — a second critique of the blanket’s scope, complementary to Bruineberg et al.
- Seth, A. K. (2015). The cybernetic Bayesian brain: From interoceptive inference to sensorimotor contingencies. In T. Metzinger & J. M. Windt (Eds.), Open MIND. MIND Group — draws the line from Ashby and Conant–Ashby to predictive processing explicitly.