Scientific modelling is a bit like playing a game of pretend. We play soldiers and pretend our sticks to be guns; we play explorers and pretend the sand to be lava. Misrepresentations, distortions, and untruths run rampant in scientific models. Such idealisations are harmless, so long as we do not forget that we have made them; so long as we do not forget that the sand is really sand and our friends really our friends.
Imagining the developing organism to be a generative model, or the brain a predictive engine can be scientifically fecund. It requires, however, a bit of suspension of disbelief. It requires the scientist to dream up a mapping between two things which are fundamentally unalike; to envision the world as though it really were a directed acyclic graph, a phase space, Shannon information measures over a Riemannian manifold. We come to talk as though these were one and the same.
Formal modelling, as with science in general, aims at uncovering causal patterns (Potochnik, Reference Potochnik2017). Philosophers of science hold scientific models to consist of interpreted structure (Weisberg, Reference Weisberg2013). Take structure to be a mathematical system; an interpretation or construal thereof relates this formal structure to a target phenomenon of interest. Model structures bear no intrinsic representation relations to targets (Nguyen & Frigg, Reference Nguyen and Frigg2021). We stipulate a relation of partial representation between model structure and target in order to put the model to work. Crucially, there will be features of any model structure that do not map onto any feature of the systems we utilise them to investigate. This makes modelling prone to undergo reification.
The mere employment of an idealisation or the application of an idealised model is not yet a reification, however; nor is the mixing of math and metaphorical language. Reification is the mismapping of formal structure onto target phenomena – or theoretical representation thereof – in a manner that leads us to misapprehend the causal structure of nature. Reification is definitionally an epistemic error.
The central move in Bruineberg et al. is to distinguish between a strictly formal reading of the Markov blanket construct, as developed in Pearl (Reference Pearl1988), and a conceptually – even metaphysically – laden notion, as wielded under the variational free energy minimization (VFEM) framework. It is this “reification of the Markov blanket construct” in the research programme centred around this second notion that they denounce.
“[M]any authors in the field,” they write, “are seemingly not aware of this process of reification, leading to the conflation of several different kinds of boundaries in the literature: Markov blankets are characterized alternatively as statistical boundaries, spatial boundaries, ontological boundaries, or autopoietic boundaries.”
One way to view this is as a multifold reification. But I think that there is a much more natural reading available, namely that the Markov blanket is simply a mathematical fixture of the VFEM framework – held as contentless formal modelling framework. It is utilised to build many different models that differ with respect to their conceptual content.
I have argued (Andrews, Reference Andrews2021) that the VFEM framework is a mathematical model structure and can do no conceptual or philosophical work on its own. The interpreted formal structure of VFEM models, however, can. Such models typically operate in an intuition-pumping capacity; they function to shape our conceptual grasp on the causal structure of target systems of interest, ranging from microphage to macrocosm (Kirchhoff, Parr, Palacios, Friston, & Kiverstein, Reference Kirchhoff, Parr, Palacios, Friston and Kiverstein2018; Kuchling, Friston, Georgiev, & Levin, Reference Kuchling, Friston, Georgiev and Levin2019; Rubin, Parr, Da Costa, & Friston, Reference Rubin, Parr, Da Costa and Friston2020). What Bruineberg et al. take to be assertions about nature are mere modelling gambits.
There is a crucial distinction to be drawn between claims made about the causal structure of nature that result from a modelling exercise and idealisations or gambits involved in the modelling procedure which may be literally untrue in reference to the model's target – and which, indeed, hold no pretence to truth (Potochnik, Reference Potochnik2017). Not all scientific modelling efforts aim to furnish such assertions, however. Some modelling work furthers our epistemic aims of latching onto causal patterns in the world by enabling us to reenvision the form that such causal architectures could take. All modelling aims at facilitating human understanding, but not all modelling aims at truth. This entails that we will have different degrees of commitment to the posits involved in our modelling efforts.
Interpreted model structure does not consist of knowledge about the natural world. The interpretation of a formal model merely maps it onto our conceptual representations of things which we hold to exist in nature. Facts, or tentative facts, are only generated when a model is put into appropriate coordination with an empirical measurement procedure. Not all scientific modelling, however, is in dialogue with data in this manner. We run into trouble when we treat modelling practices not aimed at knowledge – conceived as sufficiently true causal patterns – as generating knowledge.
Game-theoretic modelling in the social sciences or optimality modelling in biology do not on their own generate knowledge of natural systems, because they are not coordinated with the results of empirical measurement procedures. The use of VFEM models occupies a similar epistemic position. So long as we keep this context in view, the conceptual exploration facilitated by such models is harmless.
Bruineberg et al. understand VFEM models to be after claims about the causal structure of systems in the world because they implicitly take the epistemic utility of modelling to reside solely in offering truth-evaluable assertions about nature. The loose, analogical reasoning of the literature surrounding the VFEM formalism strikes them as would-be statements of fact. Were this the case, it would indeed be a reification, and an egregious one, at that. I want us to resist, however, the temptation to read the literature in this light.
Indeed, if reification were merely talk of the world as though it had formal properties, or talk of formal structure as though it had empirical or conceptual content, Bruineberg et al. would themselves be guilty of the reification fallacy, for they speak of Pearl's (Reference Pearl1988) Markov boundary as being “substantiated by the empirical literature,” and address its “scientific credibility.” The existence and use of a formal tool, however, cannot receive empirical substantiation.
Thus the takeaway of this exposition is not that the VFEM framework is unscientific or bad science, or that it does not deliver what its proponents and architects want of it, but rather that its ambitions were (like most science) far less ambitious to begin with than philosophers had hoped. The other key lesson is to caution VFEM modellers against use of overly suggestive language, lest they send the philosophers into an even deeper state of befuddlement. Such conceptual promiscuity, especially in the treatment of heavily idealised models, opens the door to reification.
Scientific modelling is a bit like playing a game of pretend. We play soldiers and pretend our sticks to be guns; we play explorers and pretend the sand to be lava. Misrepresentations, distortions, and untruths run rampant in scientific models. Such idealisations are harmless, so long as we do not forget that we have made them; so long as we do not forget that the sand is really sand and our friends really our friends.
Imagining the developing organism to be a generative model, or the brain a predictive engine can be scientifically fecund. It requires, however, a bit of suspension of disbelief. It requires the scientist to dream up a mapping between two things which are fundamentally unalike; to envision the world as though it really were a directed acyclic graph, a phase space, Shannon information measures over a Riemannian manifold. We come to talk as though these were one and the same.
Formal modelling, as with science in general, aims at uncovering causal patterns (Potochnik, Reference Potochnik2017). Philosophers of science hold scientific models to consist of interpreted structure (Weisberg, Reference Weisberg2013). Take structure to be a mathematical system; an interpretation or construal thereof relates this formal structure to a target phenomenon of interest. Model structures bear no intrinsic representation relations to targets (Nguyen & Frigg, Reference Nguyen and Frigg2021). We stipulate a relation of partial representation between model structure and target in order to put the model to work. Crucially, there will be features of any model structure that do not map onto any feature of the systems we utilise them to investigate. This makes modelling prone to undergo reification.
The mere employment of an idealisation or the application of an idealised model is not yet a reification, however; nor is the mixing of math and metaphorical language. Reification is the mismapping of formal structure onto target phenomena – or theoretical representation thereof – in a manner that leads us to misapprehend the causal structure of nature. Reification is definitionally an epistemic error.
The central move in Bruineberg et al. is to distinguish between a strictly formal reading of the Markov blanket construct, as developed in Pearl (Reference Pearl1988), and a conceptually – even metaphysically – laden notion, as wielded under the variational free energy minimization (VFEM) framework. It is this “reification of the Markov blanket construct” in the research programme centred around this second notion that they denounce.
“[M]any authors in the field,” they write, “are seemingly not aware of this process of reification, leading to the conflation of several different kinds of boundaries in the literature: Markov blankets are characterized alternatively as statistical boundaries, spatial boundaries, ontological boundaries, or autopoietic boundaries.”
One way to view this is as a multifold reification. But I think that there is a much more natural reading available, namely that the Markov blanket is simply a mathematical fixture of the VFEM framework – held as contentless formal modelling framework. It is utilised to build many different models that differ with respect to their conceptual content.
I have argued (Andrews, Reference Andrews2021) that the VFEM framework is a mathematical model structure and can do no conceptual or philosophical work on its own. The interpreted formal structure of VFEM models, however, can. Such models typically operate in an intuition-pumping capacity; they function to shape our conceptual grasp on the causal structure of target systems of interest, ranging from microphage to macrocosm (Kirchhoff, Parr, Palacios, Friston, & Kiverstein, Reference Kirchhoff, Parr, Palacios, Friston and Kiverstein2018; Kuchling, Friston, Georgiev, & Levin, Reference Kuchling, Friston, Georgiev and Levin2019; Rubin, Parr, Da Costa, & Friston, Reference Rubin, Parr, Da Costa and Friston2020). What Bruineberg et al. take to be assertions about nature are mere modelling gambits.
There is a crucial distinction to be drawn between claims made about the causal structure of nature that result from a modelling exercise and idealisations or gambits involved in the modelling procedure which may be literally untrue in reference to the model's target – and which, indeed, hold no pretence to truth (Potochnik, Reference Potochnik2017). Not all scientific modelling efforts aim to furnish such assertions, however. Some modelling work furthers our epistemic aims of latching onto causal patterns in the world by enabling us to reenvision the form that such causal architectures could take. All modelling aims at facilitating human understanding, but not all modelling aims at truth. This entails that we will have different degrees of commitment to the posits involved in our modelling efforts.
Interpreted model structure does not consist of knowledge about the natural world. The interpretation of a formal model merely maps it onto our conceptual representations of things which we hold to exist in nature. Facts, or tentative facts, are only generated when a model is put into appropriate coordination with an empirical measurement procedure. Not all scientific modelling, however, is in dialogue with data in this manner. We run into trouble when we treat modelling practices not aimed at knowledge – conceived as sufficiently true causal patterns – as generating knowledge.
Game-theoretic modelling in the social sciences or optimality modelling in biology do not on their own generate knowledge of natural systems, because they are not coordinated with the results of empirical measurement procedures. The use of VFEM models occupies a similar epistemic position. So long as we keep this context in view, the conceptual exploration facilitated by such models is harmless.
Bruineberg et al. understand VFEM models to be after claims about the causal structure of systems in the world because they implicitly take the epistemic utility of modelling to reside solely in offering truth-evaluable assertions about nature. The loose, analogical reasoning of the literature surrounding the VFEM formalism strikes them as would-be statements of fact. Were this the case, it would indeed be a reification, and an egregious one, at that. I want us to resist, however, the temptation to read the literature in this light.
Indeed, if reification were merely talk of the world as though it had formal properties, or talk of formal structure as though it had empirical or conceptual content, Bruineberg et al. would themselves be guilty of the reification fallacy, for they speak of Pearl's (Reference Pearl1988) Markov boundary as being “substantiated by the empirical literature,” and address its “scientific credibility.” The existence and use of a formal tool, however, cannot receive empirical substantiation.
Thus the takeaway of this exposition is not that the VFEM framework is unscientific or bad science, or that it does not deliver what its proponents and architects want of it, but rather that its ambitions were (like most science) far less ambitious to begin with than philosophers had hoped. The other key lesson is to caution VFEM modellers against use of overly suggestive language, lest they send the philosophers into an even deeper state of befuddlement. Such conceptual promiscuity, especially in the treatment of heavily idealised models, opens the door to reification.
Conflict of interest
None.