The Emperor’s New Markov Blankets
Our manuscript, discussing the role of Markov blankets in cognitive science, is now available as a preprint here. Work led by Jelle Bruineberg, with Krzysztof Dolega, Joe Dewhurst and myself.
- The Emperor’s New Markov Blankets, Bruineberg J., Dolega K., Dewhurst J. and Baltieri, M., A Preprint, 2020.
UPDATE: In the meantime, a heavily revised version was published on Behavioral and Brain Sciences in 2022 (available online in 2021), with a PDF inclduding our target article, various insightful commentaries and our reply to those commentaries here.
- The Emperor is Naked: Replies to the commentaries on the target article., Bruineberg, J., Dolega, K., Dewhurst, J., & Baltieri, M., Behavioral and Brain Science (2022).
- The Emperor’s New Markov Blankets, Bruineberg, J., Dolega, K., Dewhurst, J., & Baltieri, M., Behavioral and Brain Science (2021).
In this paper, we give an analysis of the development of Markov Blankets from a technical concept in Bayesian inference to a central concept within the free-energy principle/active inference framework, where they are used to define the agent/environment boundary.
We argue that it is important to distinguish the original instrumental use of Markov blankets (which we call for short ‘Pearl blankets’) from their more recent, seemingly realist use in the FEP/active inference framework (which we call ‘Friston blankets’).
The latter have been used to draw metaphysical or ontological conclusions, which we think require additional metaphysical premises, and cannot be justified just on the basis of the formal success of the original Markov blanket construct (see figure below).
We believe that the two constructs have been conflated in the literature, which has caused confusion when Markov blankets are supposedly used to define the ontological boundary of a system, e.g. mind/world or organism/environment.
It would therefore be beneficial to keep these two constructs (Pearl blankets and Friston blankets) distinct, using the former for purely instrumental interpretations of statistical models, and the latter (with additional premises) to draw stronger metaphysical conclusions.
Our work follows some recent relevant contributions including
- Mel Andrews’ excellent recent paper “The Math is not the Territory”
- Menary & Gillett’s “Are Markov blankets real and does it matter?”
- Fernando Rosas et al.’s “Causal Blankets: Theory and algorithmic framework”
- Martin Biehl et al.’s “A technical critique of the free energy principle as presented in “Life as we know it” and related works”
(The last two, in particular, provide some of the technical background required to understand and evaluate whether Friston blankets are a well defined concept, one that can be used in the future to support the metaphysical claims in the FEP literature.)
