In this commentary we argue that, from the point of view of Bayesian cognitive science, concerns about the reification of Markov blankets are misplaced. We assume that scientific theories represent reality in the same way in which we have argued (Kiefer & Hohwy, Reference Kiefer, Hohwy and Kirchhoff2017, Reference Kiefer, Hohwy, Robins, Symons and Calvo2019) that organisms' internal generative models represent their external worlds: by way of (exploitable) structural similarity. Mathematically articulated theories or models set up relations among their variables, which, when successful, mimic the relations that obtain among elements of the target system (Cummins, Reference Cummins1991). In asserting such theories, we impute these very structural properties to the world.
This view is fundamentally a structural-realist one (Ladyman, Reference Ladyman1998; Ladyman & Ross, Reference Ladyman and Ross2007), and is in a sense nothing new, but whether novel or not it is adequate to the task of addressing the present criticism. To start, this perspective on the semantics of scientific theories completely defuses the concern that projection of the content of a mathematically expressed theory (such as the free energy principle [FEP]) onto the physical world involves a category-mistake or fallacy. The “substance” of the representational medium (whether collections of abstract mathematical symbols or bits of pasta) makes no difference apart from its expressive adequacy, since it is the form and not the substance of the target domain that the theory attempts to capture in virtue of its form.
A traditional argument against instrumentalism (Putnam, Reference Putnam1975) has it that theories are predictive only to the extent that they are true. Structuralist representationalism allows us to nuance this argument: We may expect observational adequacy to the degree that the structure of the theory matches that of the generator of observations. A Bayesian might point out that the confidence we (ought to) place in a theory also scales with its predictive accuracy (as well as its prior plausibility). By adopting a Bayesian attitude towards scientific theories, we can then dispense with a categorical distinction between realism and instrumentalism, while taking on board the epistemic humility that makes the latter appealing.
The preceding remarks do not on their own, of course, justify a realist attitude towards Markov blankets in living systems or elsewhere – this would depend on the empirical adequacy of such Markovian descriptions. They do however suggest that there is nothing methodologically flawed in the practice of imputing formally characterized structures (such as structures of conditional independence in Bayesian networks) to the real world.
In the remainder we rebut three arguments given in the target article that purport to establish the contrary.
The first argument poses an analogy between Markov blankets and contour lines on cartographic maps. I expect to observe rivers and mountains when I navigate by a map, but pace the authors, I also expect to experience elevation and other aspects of the terrain represented by contour lines. To mistake contour lines as such for features of the terrain would indeed be a radical mistake, but so would expecting the river to be ink-blue. We do not make these mistakes in practice any more than proponents of the FEP suppose on the basis of their diagrams that the sensory epithelia of organisms consist in labelled circles with black outlines.
The moral is that contour lines contribute to the same fundamentally spatial (indeed structural-similarity-based) representation as other elements of the map, though in a slightly more abstract and conventional way. There may be more reasonable concerns about reification, for example that graphical models cut the world artificially into discrete, repeatable event-types, but this worry would impugn the use of such abstractions to represent causal structure (such as smoking's causing cancer) quite generally, and so is not properly directed against the FEP literature.
In fact, the authors do argue for a blanket instrumentalism with respect to Bayesian networks, which brings us to the second argument: That the choice of model for a given system depends in part on extrinsic circumstances like data availability or the interests of the scientist. But the intrusion of this pragmatic element means merely that there exist many sets of conditional dependencies, some more interesting than others, which does not impugn a realist attitude towards any one of those sets.
The authors further suggest that the Markov blanket formalism does little work in delineating organismic boundaries if we must already have selected a model in order to consider its blanket. It is unclear to us how this epistemic point could count against realism, but in any case the plurality of Markov blankets present in any system of interest has long been admitted by FEP theorists (cf. Friston et al., Reference Friston, Fagerholm, Zarghami, Parr, Hipólito, Magrou and Razi2021; Hohwy, Reference Hohwy2016), and in practice the FEP is interesting not as a tool for distinguishing organisms from their environments, which we can do well enough without it, but for its formal account of how systems act to maintain the integrity of their boundaries, however initially identified.
The final argument we consider suggests that the literature around the FEP stratifies into two distinct projects, the respectable empirical one of using Markov blankets to characterize aspects of organisms' cognitive models of their environments (“inference with a model”), and the metaphysically ambitious one of using Markov blankets to characterize organisms themselves and their boundaries with respect to their environments (“inference within a model”). Here, it is sufficient to point out that these projects are not, actually, fundamentally distinct in kind. The explanatory targets in both cases are real features of organisms (their cognitive models, in the first case, and their sensorimotor boundaries, in the second). What may be controversial in the case of the FEP is the idea that the entire organism (as opposed to some construct in its brain) may be regarded as a “model,” but this is not the ground on which the authors stake their claim.
In conclusion, we have seen no reason, either on the basis of its general mathematical character or on the basis of its particular modes of application, to suspect that the Markov blanket formalism has been used by FEP theorists to commit fallacies of reification.
In this commentary we argue that, from the point of view of Bayesian cognitive science, concerns about the reification of Markov blankets are misplaced. We assume that scientific theories represent reality in the same way in which we have argued (Kiefer & Hohwy, Reference Kiefer, Hohwy and Kirchhoff2017, Reference Kiefer, Hohwy, Robins, Symons and Calvo2019) that organisms' internal generative models represent their external worlds: by way of (exploitable) structural similarity. Mathematically articulated theories or models set up relations among their variables, which, when successful, mimic the relations that obtain among elements of the target system (Cummins, Reference Cummins1991). In asserting such theories, we impute these very structural properties to the world.
This view is fundamentally a structural-realist one (Ladyman, Reference Ladyman1998; Ladyman & Ross, Reference Ladyman and Ross2007), and is in a sense nothing new, but whether novel or not it is adequate to the task of addressing the present criticism. To start, this perspective on the semantics of scientific theories completely defuses the concern that projection of the content of a mathematically expressed theory (such as the free energy principle [FEP]) onto the physical world involves a category-mistake or fallacy. The “substance” of the representational medium (whether collections of abstract mathematical symbols or bits of pasta) makes no difference apart from its expressive adequacy, since it is the form and not the substance of the target domain that the theory attempts to capture in virtue of its form.
A traditional argument against instrumentalism (Putnam, Reference Putnam1975) has it that theories are predictive only to the extent that they are true. Structuralist representationalism allows us to nuance this argument: We may expect observational adequacy to the degree that the structure of the theory matches that of the generator of observations. A Bayesian might point out that the confidence we (ought to) place in a theory also scales with its predictive accuracy (as well as its prior plausibility). By adopting a Bayesian attitude towards scientific theories, we can then dispense with a categorical distinction between realism and instrumentalism, while taking on board the epistemic humility that makes the latter appealing.
The preceding remarks do not on their own, of course, justify a realist attitude towards Markov blankets in living systems or elsewhere – this would depend on the empirical adequacy of such Markovian descriptions. They do however suggest that there is nothing methodologically flawed in the practice of imputing formally characterized structures (such as structures of conditional independence in Bayesian networks) to the real world.
In the remainder we rebut three arguments given in the target article that purport to establish the contrary.
The first argument poses an analogy between Markov blankets and contour lines on cartographic maps. I expect to observe rivers and mountains when I navigate by a map, but pace the authors, I also expect to experience elevation and other aspects of the terrain represented by contour lines. To mistake contour lines as such for features of the terrain would indeed be a radical mistake, but so would expecting the river to be ink-blue. We do not make these mistakes in practice any more than proponents of the FEP suppose on the basis of their diagrams that the sensory epithelia of organisms consist in labelled circles with black outlines.
The moral is that contour lines contribute to the same fundamentally spatial (indeed structural-similarity-based) representation as other elements of the map, though in a slightly more abstract and conventional way. There may be more reasonable concerns about reification, for example that graphical models cut the world artificially into discrete, repeatable event-types, but this worry would impugn the use of such abstractions to represent causal structure (such as smoking's causing cancer) quite generally, and so is not properly directed against the FEP literature.
In fact, the authors do argue for a blanket instrumentalism with respect to Bayesian networks, which brings us to the second argument: That the choice of model for a given system depends in part on extrinsic circumstances like data availability or the interests of the scientist. But the intrusion of this pragmatic element means merely that there exist many sets of conditional dependencies, some more interesting than others, which does not impugn a realist attitude towards any one of those sets.
The authors further suggest that the Markov blanket formalism does little work in delineating organismic boundaries if we must already have selected a model in order to consider its blanket. It is unclear to us how this epistemic point could count against realism, but in any case the plurality of Markov blankets present in any system of interest has long been admitted by FEP theorists (cf. Friston et al., Reference Friston, Fagerholm, Zarghami, Parr, Hipólito, Magrou and Razi2021; Hohwy, Reference Hohwy2016), and in practice the FEP is interesting not as a tool for distinguishing organisms from their environments, which we can do well enough without it, but for its formal account of how systems act to maintain the integrity of their boundaries, however initially identified.
The final argument we consider suggests that the literature around the FEP stratifies into two distinct projects, the respectable empirical one of using Markov blankets to characterize aspects of organisms' cognitive models of their environments (“inference with a model”), and the metaphysically ambitious one of using Markov blankets to characterize organisms themselves and their boundaries with respect to their environments (“inference within a model”). Here, it is sufficient to point out that these projects are not, actually, fundamentally distinct in kind. The explanatory targets in both cases are real features of organisms (their cognitive models, in the first case, and their sensorimotor boundaries, in the second). What may be controversial in the case of the FEP is the idea that the entire organism (as opposed to some construct in its brain) may be regarded as a “model,” but this is not the ground on which the authors stake their claim.
In conclusion, we have seen no reason, either on the basis of its general mathematical character or on the basis of its particular modes of application, to suspect that the Markov blanket formalism has been used by FEP theorists to commit fallacies of reification.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Conflict of interest
None.