IS THE FREE ENERGY PRINCIPLE FOR REAL?

‘Intelligence without ambition’, said Salvador Dali, ‘is like a bird without wings’. Certainly, there exists an ever-increasing number of researchers who take the free energy principle (FEP) and its mathematical formalism to provide a vital new apparatus for fulfilling even their grandest intellectual ambitions. And how grand these ambitions are! The FEP was initially invoked as a theoretical framework in attempts to provide a unified theory of brain function (Friston [2010]). More recently, its proponents have sought to use it to explain how all biological organisms self-organize so as to resist the natural tendency towards disorder (for discussion, see Constant [2021]). As Jakob Hohwy ([2016], p. 729) notes, not only does the FEP aim ‘to explain everything about the mind’, but it also does so from the vantage point of a framework that takes itself to have, in Shaun Raviv’s ([2018]) words, ‘identified nothing less than the organizing principle of all life’.

It has been noticed in popular venues that the FEP, possessing as it does enough ambition to make Don Quixote blush, is, again to quote Raviv ([2018]), ‘an idea every bit as expansive as Darwin’s theory of natural selection’. It is of small wonder, then, that many researchers have interpreted its purportedly prodigious explanatory scope as a clear theoretical red flag (Froese and Ikegami [2013]; Klein [2018]; for discussion, see also Hohwy [2015]). After all, as its proponents readily admit, it will initially strike many as ‘preposterous’ that all cognitive activity somehow adheres to one simple mathematical principle (Hohwy [2015], p. 1).

To broach the question of whether the FEP can fulfil its dizzying explanatory ambitions, it is first crucial to understand how exactly the FEP’s mathematical models are supposed to describe the kinds of phenomena to which they are meant to apply. How, for example, do FEP models of neural activity provide mathematical descriptions of said neural activity? How does such mathematical modelling map onto the target phenomena? And here a philosophical controversy has recently reared its head, in the form of the venerable and ongoing debate between scientific realists and instrumentalists.

FEP researchers are presently divided as to whether we should construe the mathematical models yielded by the FEP framework as offering something like accurate or approximately accurate descriptions of their target phenomena (for recent discussion, see Andrews [2021]). Some say yes, some say no. The naysayers endorse an instrumentalist construal of the FEP, according to which its models amount merely to heuristic tools for generating, with increasing reliability, observational predictions. FEP models, on this analysis, are not in the business of making statements or providing accurate descriptions of their target phenomena. FEP models are really just useful fictions (and, indeed, controversy persists over whether FEP models offer all that much in terms of explanatory traction as compared to those yielded by other frameworks; see, for example, Colombo and Wright [2021]).

There has been a surge of papers defending precisely such an instrumentalist construal of the FEP. In our recent article, we provide a response on behalf of the realist. We argue that realism about FEP models remains a live and tenable option. Although we do not claim that all FEP models should be construed as statements or representations of their target phenomena, we hold that the idea that many of its models do exactly this should be taken seriously. In attempting to secure our realist conclusions—to advance our case that some FEP models really are in the business of accurately describing the phenomena they model—we identify what we take to be a pervasive mistake in the contemporary FEP literature, which we refer to as the ‘literalist fallacy’. We take it that many instrumentalists make this literalist fallacy in their dismissal of realist interpretations of FEP models.

Before explaining the literalist fallacy, a short description of the FEP is required (for a more technical description, see our article). Simply put, the FEP offers an information-theoretic solution of an inference problem (key assumption) that every living organism, from cells to humans, must overcome:  inferring the probability of the causes of one’s sensory observations conditioned on prior beliefs about such causes. This much looks crucial for survival. The game of life (under the FEP) is thus to reduce the inherent uncertainty about (unobservable) causes, given their (observable) effects by continuously updating Bayesian ‘beliefs’ about the causes, given the effects, that is, the probability of the cause given the effect. The FEP offers the insight that what is called ‘variational free energy’ can be used as a mathematical proxy that functions as a bound on the ‘surprise’ inherent in sensory observations—such that minimizing free energy is equivalent to reducing uncertainty about causes, thereby precluding surprising sensory effects that would be uncharacteristic of the kind of life in question (for a simplified, philosophically friendly introduction to the FEP, see Mann et al. [2022]).

What, then, do we mean by the literalist fallacy? In the context of the FEP, we claim that the literalist fallacy occurs when the truth of instrumentalism about FEP models is inferred on the basis of an overly literal and demanding notion of what a realist must claim about how the mathematics of its models map onto their target phenomena. We point to examples of this, namely, where realists about the FEP are taken by their instrumentalist opponents as being committed to organisms having to somehow literally embody or operate in direct accordance with the mathematical structure of the FEP models invoked to model them. The literalist fallacy, then, stems from a mistaken and impoverished idea of what scientific realism really amounts to. Indeed, although our article concerns only the FEP framework, we take something akin to the literalist fallacy to pervade much of the debate between scientific realists and instrumentalists.

Ultimately, we suspect that many instances of the literalist fallacy stem from a misunderstanding of how scientific realists seek to account for abstraction and idealization in contemporary scientific modelling enterprises. One such instance of this, we submit, is in (Ramstead et al. [2020]). While investigating whether FEP models amount to more than predictive instruments, they claim that ‘instrumentalist accounts in the philosophy of science suggest that scientific models are useful fictions: they are not literally true, but “true enough”, or good enough to make useful predictions about, and act upon, the world’ ([2020], p. 4). But all scientific realists will here insist that models can be literally partially true of their target phenomena. In fact, that a model contains sufficient truth to have predictive prowess seems to be exactly in keeping with paradigm statements of scientific realism (for philosophical discussion of how the distinction between scientific realism and instrumentalism tends to be carved up by philosophers, see Haack [2004]; see also Williamson [2017]; Massimi [2021]).

In order to expose the faulty assumptions leading to the literalist fallacy, and to demonstrate that realism about the FEP remains a tenable option, our BJPS article considers a host of ways to interpret FEP models under realist construals. We note that models are now being used productively to gain insight into a plethora of different phenomena: decision-making under uncertainty (Friston et al. [2012]), optimal control (Çatal et al. [unpublished]), psychopathology (Schwartenbeck et al. [2015]), active scene construction (Mirza et al. [2016]), electrophysiological responses (Friston et al. [2017]), and so on (for a detailed overview of different applications of FEP models, see Da Costa et al. [2020]). Adopting active inference models of eye-tracking as a concrete case study, we additionally claim that the effectiveness of such models constitutes a body of prima facie evidence for realism.

We refrain here from discounting or downplaying the deeply vexed philosophical questions that emerge when we pose Eugene Wigner’s ([1960]) famous questions as to why mathematics seems to have an ‘unreasonable effectiveness’ in guiding scientific endeavour. That said, avoiding the literalist fallacy is crucial when properly evaluating realist answers about the explanatory capacity of the FEP’s mathematical framework. In doing so, we should be clear that scientific realists rarely strive to provide ‘complete, non-distorted, perfectly accurate representations’, as Michael Weisberg ([2007], p. 657) correctly notes, and that, given this, short-term realist practice usually ‘involves the willful introduction of distortion’ in exchange for long-term gain and increasing representational accuracy. In assessing the various ways that we might adopt a realist interpretation of FEP models, our article considers at length how such models incorporate complex idealizations and abstractions. We also consider whether FEP models are best construed as making Galilean abstractions.

As the FEP’s research potential continues to be realized (with the active inference scheme being invoked as a process theory), questions about the applicability of its formalism become increasingly pressing (Colombo and Wright [2021]; Ramstead et al. [unpublished]; Van Es and Hipólito [unpublished]). Avoiding the literalist fallacy will help to ensure that we do not rule out one potentially promising set of answers.