1. Introduction
The Free Energy Principle (FEP) inspires one of the most comprehensive frameworks for the study of complex adaptive systems. Originated by Karl Friston and colleagues, it spans active research in several life science fields under the overarching principle of internal free energy minimization. For this reason, the FEP can be counted among contemporary attempts to understand complex systems through variational principles—see e.g., [
1,
2,
3,
4] for a short review. However, the FEP seems to encompass a broader picture because of its alleged ability to include more restricted approaches in neurosciences, including the mind-brain problem ([
5], p. 136). What is more, some scientific commentators claim it might hold the key to artificial intelligence [
6].
Throughout the last two decades, FEP formalism has been presented with technical aspects that may prove threatening for the uninitiated. Admittedly, changes to notation and symbols can misguide readers, hence FEP papers typically include tables with mathematical definitions of key concepts and figures that help visualize the underlying schema. Nevertheless, the theory increasingly appears as a promising and encompassing framework to study living systems, both dia- and syn-chronically, and has already been considered in the literature as a grand unifying theory (GUT) ([
7], pp. 5–7) for the understanding of evolution, sentience and consciousness, as well as of human cognition and the acquisition of culture [
8]. Some have even argued that the FEP introduces an unresolved tension between the secluded organism and its necessary openness and interaction with the environment, which finds its philosophical home in Hegel’s dialectics [
9].
During the last decade, criticism of the FEP has emerged, see e.g., [
10]. They cannot but be welcomed inasmuch as the FEP’s initial assumptions and philosophical consequences require clarifications and discussion of their drawbacks. Recently, however, the FEP has been attacked on more general grounds—targeted for its distinction as a possible GUT. Whereas such criticism might be partially justified in terms of the FEP’s failure to achieve its goal as it stands, it might also go too far and ruin the main thrust of scientific reduction. Hence the existence of apparently irreducible models to explain, for instance, the activity of mesocorticolimbic dopaminergic systems [
7] need not necessarily count as evidence against the FEP’s aspirations; on the contrary, it should stir up better explanations through proper modeling within the FEP framework.
This paper assumes the FEP’s implicit claim of epistemic reduction in its understanding of complex adaptive systems and concentrates on reviewing its explanatory power. Beyond technical criticisms regarding the emergence of Bayesian inference via Markov blankets and the FEP’s universality [
11], Colombo and Wright [
12] tried to clarify how the FEP relates to the two most prominent theoretical approaches to life science phenomena, namely organicism and mechanism. Notwithstanding, Friston and colleagues have recently published a discussion of the FEP’s compatibility with different philosophical perspectives, ranging from monism to dualism [
13], and explicitly adopting what they dub Markovian Monism (MM). This move naturally demands careful review of the philosophical assumptions behind the FEP’s main concepts and tenets, as well as its links (or lack thereof) to key philosophical concepts like representation, supervenience and individuation.
Should the FEP’s newly minted ontology be understood as a response to general criticism of the FEP as a GUT? The present work maintains both that foregoing critiques—especially those by Colombo and Wright [
12]—can be rejected on grounds other than MM and that the latter is a misleading basis for the FEP framework due to its circular reasoning. After introducing the FEP’s main concepts and assumptions in non-technical jargon in
Section 2, this paper endeavors to (1) distinguish between misplaced and valuable criticisms in Colombo and Wright’s review of the FEP by defending its role as a scientific theory (
Section 3 and
Section 4) and (2) assess the value of MM as an interpretation of the FEP.
Section 5 distils the most relevant ideas in MM,
Section 6 makes explicit some of MM’s assumptions regarding key philosophical concepts, and
Section 7 begs the fundamental question of extant individual systems in nature. Ultimately, the FEP may withstand criticism inasmuch as it is not committed to MM for two main reasons: First, as the FEP’s philosophical interpretation, MM obtains what it assumes, i.e., a dual information geometry that allegedly explains epistemic duality because of its prior, initial dual assumptions. Second, such circular reasoning stems from the FEP’s implicit reliance on a non-fundamental, relative system-environment separation. The conclusions herein explain why this double circularity proves harmless for the FEP inasmuch as it remains a principle limited to science, providing epistemic advancement to life sciences. Whereas the FEP as a scientific theory can be based on such circularities, the FEP’s philosophical interpretations, like MM, cannot. Science needs philosophical assumptions through which its theories can be linked with observations but does not need to justify them if they remain open to philosophical criticism. However, philosophical interpretations of scientific theories that aim at ultimate foundations may not contain unresolved circular reasoning. That is why MM is a questionable philosophical backdrop for the FEP.
3. The (Misplaced) Criticism of the FEP as a GUT
Before we turn to the philosophical questions raised by the FEP framework and its characterization of complex adaptive systems, it is worth devoting some space to Colombo and Wright’s critique of the FEP’s attempt to reach GUT status [
7]. These authors draw on a purportedly central case, the activity of mesocorticolimbic dopaminergic (DA) systems, in which the different explanatory models for DA systems, namely anhedonia, incentive salience, and reward prediction error hypotheses, seem to vindicate explanatory pluralism and demonstrate that scientific progress in the cognitive sciences is unlikely to be associated with a single overarching GUT. Is such an inference justified?
In cognitive science, the FEP framework can be interpreted as a form of error correction via Bayesian dynamics of previous beliefs, instantiated by a bidirectional cascade of cortical processing, see e.g., [
28,
35,
36,
37]. In fact, human observers behave as optimal Bayesian observers in many ways ([
38], p. 712). Each level’s basic schema sets out that, “feed-forward connections convey information about the difference between what was expected and what actually obtained—i.e., prediction error—while feedback connections convey predictions from higher processing stages to suppress prediction errors at lower levels” ([
7], p. 5). With its conceptual simplicity and powerful formalism, the free energy theory lends the ultimate rationale for Bayesian cycles of perception and action, attaining the goal of unification and scientific reduction of other higher-level explanatory principles (e.g., those utilized in psychology).
Nevertheless, Colombo and Wright [
7] complain about what, for the time being, is only a would-be GUT. Currently, the FEP suggests the possibility of genuine future inter-theoretic reductions of higher-level theories to the FEP framework. However, the three abovementioned models regarding the activity of mesocorticolimbic DA systems remain irreducible and, contrary to GUT expectations, show how actual scientific practice vindicates explanatory pluralism. For instance, terms like reward and value are deemed irreducible via mathematical ‘absorption’ in favor of prior beliefs. Explanatory pluralism rejects the suggestions that, for any phenomenon, there will always be exactly one single, complete, comprehensive explanation based on a single set of fundamental principles. It assumes that scientific theories co-evolve and mutually influence one another without lower-level theories supplanting higher-level theories and hypotheses. It is precisely this inter-theory competition and selection pressure that accelerates scientific progress. Hence, “[p]rogress in neuroscience is ill-served by fervently advancing a single GUT of mind/brain that attempts to solve all problems. Rather, it is more productive to focus experimental and theoretical research on some problems, and to generate a plurality of solutions that compete as local explanations and narrowly-conceived hypotheses” ([
7], pp. 6–11).
My argument is that, within this scientific context, the foregoing remarks are misplaced and unfair to the FEP framework. Not only does some of this critique turn out to be controversial, like the alleged irreducibility of reward and value to the role of priors, but, more importantly, denying the legitimacy of attempts at GUTs—even if only within neurosciences—entails a narrowly-conceived view of science. There might well be simplified, higher-level models to tackle specific practical problems, but scientific activity cannot dispense with its search for greater unification without putting meaning at risk or, worse, becoming sheer empiricism. True, Colombo and Wright [
7] point out relevant, specific problems that the FEP framework needs to tackle in order to keep up with its aspirations to become a GUT. Yet, FEP does not claim at this stage to be a GUT; to date, it provides theoretical neuroscience and life disciplines with a new paradigm that can, in principle, encompass otherwise partial frameworks and shallower explanatory logics, much as evolution became the universally shared framework for biology. And, of course, the FEP may become ultimately unsuccessful.
The existence of hitherto irreducible models is not a drawback, but rather stimulates the search for more complete scientific explanations, which is a deeper stimulus than underscoring the mutual benefit of theories’ pluralistic co-evolution ([
7], pp. 10–11). Colombo and Wright do not justify their pluralistic approach with such difficulties more than difficulties in explaining specific transitional states could justify finishing off the evolutionary framework. Moreover, explanatory pluralism as a last-ditch epistemic strategy smacks of an attempt at unassailability and obtains little by dismissing the epistemic core of scientific explanation. Does this mean that a scientific GUT, like the FEP aims to be, is bound to be the ultimate explanation? Not necessarily since GUTs may urgently require philosophical clarifications and/or interpretations. I will address this issue in
Section 5,
Section 6 and
Section 7 when discussing MM, but let us first identify more profound critiques of the FEP in the foregoing authors.
4. Valuable Critiques of the FEP
Fortunately, Colombo and Wright have other relevant and timely critiques of the FEP that deserve closer inspection [
12]. They ask for the detailed philosophical review in
Section 5,
Section 6 and
Section 7 and require clarification of the FEP’s epistemic status: “FEP’s epistemic status remains opaque, along with its exact role in biological and neuroscientific theorizing. Conspiring against its accessibility are the varying formalisms and formulations of FEP, the changing scope of application, reliance on undefined terms and stipulative definitions, and the lack of clarity in the logical structure of the reasoning leading to FEP”.
Colombo and Wright initially praise the FEP framework as a powerful attempt at blending biology and information that might also illuminate the continuity between life and mind, as the FEP applies to any biological system. Nevertheless, concerns arise due to the FEP’s apparent silence on the nervous system’s biophysical reality and its implicit commitment to some form of cognitivism, where cognition is taken for granted within a functional scheme. However, functional analyses lack explanatory power as they are “sketches of mechanisms, in which some structural aspects of a mechanistic explanation are omitted. Once the missing aspects are filled in, a functional analysis turns into a full-blown mechanistic explanation” ([
39], p. 283). Accordingly, mechanists play down the FEP’s explanatory role because of its lack of specific analysis, including the biophysical details to localize each operation with its respective component part: “Phenomena studied in the life sciences should be explained by appealing to the component parts and operations of mechanisms, where a mechanism is a spatiotemporally-organized composite system producing a phenomenon” [
12]; see e.g., [
40,
41,
42] for an improved mechanistic explanation of complex adaptive systems. Briefly stated, for mechanists, efficiency and structure determine form and function. However, obviously this perspective is itself highly controversial. (Functional explanations allow emergent functions to be realizable in multiple ways in complex adaptive systems, obviously supported by microphysical states, but not necessarily reducible to the latter’s dynamics. In this sense, functional explanations may provide functions with an ontological status beyond epiphenomenalism, which reductive mechanists may see as a drawback.)
However, to avoid getting ahead of myself, here I will focus on criticism from the organicist perspective, which raises serious doubts about the ability of physics to adequately represent organisms and their behaviors. On the one hand, organicists accuse FEP theorists of too quickly ascribing FEP framework properties and tools to organisms. Do free energy, surprise, and optimization exist as biological properties in complex adaptive systems? One could bring such concerns to the general problem of representation in scientific theories, which affects topics beyond the FEP framework, although obviously exceeds the scope of this paper.
On the other hand, organicists heavily criticize the FEP’s crucial assumption about ergodicity over the system’s span of existence. This assumption might lead towards insufficient characterizations of the organism’s phenotype and defining properties. Admittedly, FEP theorists always introduce the caveat of applying formalism to well-defined, steady phases of the system’s life because, by definition, the FEP framework cannot deal with processes like death. Ergodicity “only holds over certain temporal scales for real organisms that are on a trajectory from birth to death” ([
5], sec. S1). Yet, such statements seem to fall into a sort of tautology, as the FEP framework works for ergodic phases and is unable to predict more complicated life phenomena related to phase transitions. In other words, overall, life behaves non-ergodically for organicists. Organisms “live” in extended critical phase transitions [
43] or on the edge of order and chaos in the region of criticality [
44,
45], which disavows homeostatic stability as the core feature of living systems. (One anonymous reviewer pointed out that “ergodicity is not a problem for the FEP framework because the FEP framework only assumes local ergodicity: relative to the parts that make it up, each whole looks stationary”. Obviously, such a view is highly problematic when degrees of freedom are redefined and parts of the alleged whole are redefined—what is related to the individuation problem, more on this in
Section 7. Certainly, even if life might be globally non-ergodic, it is locally ergodic enough to, e.g., be considered as having locally stable properties. Nevertheless, the problem arises at critical transition points. To follow up with an analogy suggested by the reviewer, Earth might be flat enough to build skyscrapers, but not to assume that interoceanic flights follow a straight line. To put it plainly, one must be aware of accumulated error in dealing with living systems as a temporal line of stable cycles. Such error may dramatically increase when it comes to a critical point.)
If that is the case, proper description of life might be not only non-computable in a finite number of steps, but also non-algorithmic. Because of their intrinsic historicity, organisms need not possess general characteristics that allow for complete mathematically invariant representations. Up to what point can one speak of biological symmetries that have to be preserved? Certainly, one is allowed to tackle such symmetries as a good approximation to life characteristics—see e.g., [
46]—but breaks in symmetry that redefine the relationships between system and environment continually occur and some contingency beyond general principles might be crucial for understanding organisms. Organicists reject mere adaptationist and selectionist perspectives because of their insufficient explanation of the autonomy of living beings for regulating their processes in relation to environmental conditions [
12]. On the contrary, FEP advocates rely on the power of formalism to incorporate any regulation whatsoever. Such disagreement becomes transparent in the conceptual differences regarding what counts as “surprising” for an organism. Whilst organicists stress that surprising events need not always be maladaptive for organisms, Friston and colleagues answer by nuancing the term “surprise”, making it contingent upon the context in which the FEP framework works and the “temporal depth” and “epistemic affordances” of different kinds of sentient systems [
13].
Importantly, the FEP framework relies on representations of the external world via a generalized inferential picture of cognition. Obviously, the system needs to start sampling the environment according to certain prior beliefs—equivalent to setting initial conditions—that allegedly recapitulate environmental patterns, but do not in themselves enable efficient inference; the FEP needs to be invoked for that. However, it might also happen that, even if formalism by construction converges on a course of action, trajectory or policy, it fails to reproduce the actual system behavior because of unforeseeable changes in its phase space. Last but not least, since the FEP framework admits the duality of information geometries in its probability distributions —intrinsic (for physical states) and extrinsic (referring to belief states)—how should such probabilities be consistently interpreted? Are interpretations of probabilities as physical propensities compatible with interpretations of probabilities as cognitive inferences at all levels? If, as Colombo and Wright argue [
12], Friston interprets probabilities involved in the FEP as objective features of systems [
17], a clash seems to emerge between (physical) propensities as causal tendencies that should be asymmetric, like causal relationships, and epistemic probabilities (cognitive inferences) that should not. Conditional probabilities can be reversed—in this sense, they are symmetric, even if reversed conditional probability might be very different from the initial probability. However, cause and effect relationships cannot in general be reversed. Convergence of the information geometry’s dual aspects onto a single objective probability may thus not be that straightforward. (The degree of identification between the two aspects in information geometry is far from clear, as Kiefer has recently pointed out [
47]. Obviously, if one-to-one correspondence is relaxed, MM is in trouble as an ontology for the FEP. As a matter of fact, the identification of physical and cognitive aspects in a single physical concept is becoming fashionable in some theories about consciousness. The most well-known example is Integrated Information Theory (IIT)—see e.g., [
3], where phenomenological axioms are straightforwardly translated into mathematical axioms. Yet, such a procedure is defective without further clarification [
48].)
7. What Is an Individual System?
One of organicists’ deepest criticisms of the FEP, as hinted at in
Section 4, amounts to the degree of specificity of physical systems. “If historical considerations and lineage matter to understanding organisms and their dynamics, then biological systems should be represented as ‘specific’ and their trajectories as ‘generic.’ Instead … free-energy theorists get it backward: physical systems are ‘generic,’ while their trajectories ‘specific’” ([
12,
43], Chapter 7). Behind that critique, one may divine a certain mistrust for understanding individual organisms through the FEP alone. Moreover, if living systems are not just optimization processes, but rather “extended critical transitions, always transient toward a continually renewed structure” ([
43], p. 162), their individuality may turn out to be unique in the sense of the impossibility of wholly explaining them away by general principles.
Remarkably enough, FEP theorists start off with the observation that some systems maintain their physical integrity, displaying adaptive behavior in a changing environment. Systems are featured as remaining in a non-equilibrium steady state, within a relatively narrow region of all possible states of their initial phase space. If that is the case, Markov blankets can be deemed objective features of nature, separating biological systems’ internal states from those external to them [
17]. Then, “by acting on the environment to minimize the free energy of their sensory samples, biological systems would avoid surprising sensory states. If they avoid surprising sensory states, biological systems may attain a homeostatic state; and by selecting actions that attain homeostatic states, biological systems will thereby behave adaptively and preserve their physical integrity” [
12].
However, such “generalized homeostasis” ([
13], n. 7) rests upon delicate, non-fundamental, stipulations—e.g., contingent upon a concrete range of energy exchange for some timespan (the duration over which the system exists) between the system and the environment. Only within that regime does it make sense to a priori define the relevant degrees of freedom for the system, including a specific Markov blanket parametrizing the system’s interaction with the environment. Yet, is it possible to a priori define the relevant degrees of freedom or states for complex adaptive systems? Such systems can continually increase the diversity of what happens next [
45]. How might we define the system’s relevant degrees of freedom in fundamental terms when complex adaptive systems’ autopoiesis remains controversial [
56]? Even with an effort in the FEP literature to link priors to phenotypes [
49] in the FEP’s framework, existence just means the attainment of a steady state in a system without explaining its evolutionary history. However, such a perspective is flawed as it cannot encompass a system’s changing phase space or non-steady probability densities related to the system’s states, both of which might prove essential in understanding life phenomena.
In other words, the FEP suffers from circularity because it a priori assumes conditions that are to be maintained. No novelty can emerge without further assumptions within this model. For Friston and colleagues, “A particle or person is never ‘off’ their manifold—they just occupy states that are more or less likely, given the kind of thing they are” [
13]. Still, life phenomena deserve further explanation, namely why does identity remain in an ever-changing universe? Is identity merely epistemic or contextually ontic? If it is the latter, why do such contexts allow for the emergence of individual systems? The FEP framework simply takes as given that particles, persons or any system whatsoever are “the kind of thing they are”. That may be enough for science, but it is far from enough for philosophy.
In short, the contextuality of system/environment separation always lurks behind any would-be fundamental theory of life and sentience, and the FEP framework is not immune to that. The FEP deserves credit for showing how “action and perception look as if they are minimizing a particular entropy” [
13]. Yet, it is only too well-known that physics still faces a fragmentary landscape regarding the application of variational principles to free energy minimization or entropy production. What is more, a pragmatic approach seems the only way forward as far as “the empirical and numerical evidence appears to suggest that there is no universal entropy production functional that is maximized in all problems” ([
57], p. 19).
Nonetheless, in as far as its limitations are acknowledged, such circularities need not result in a drawback for the theory. The FEP might be circular, but, in moving around a circle whose foundations are accepted, it provides epistemic gain through a unified principle featuring homeostatic-like phases in living systems’ complex phase space. There, any system characterized as possessing a Markov blanket “can be cast as performing some elemental form of inference—and possessing an implicit generative model” [
13]. As a good ally to complexity sciences, FEP formalism introduces a specific frame in which meaning itself might emerge as intimately related to systems’ survival and natural selection: “Events that happen to an organism mean something to that organism if those events affect its well-being or reproductive abilities. In short, the meaning of an event is what tells one how to respond to it … This focus on fitness is one way I can make sense of the notion of meaning and apply it to biological information-processing systems” ([
58], p. 184). Even so, the FEP’s circularities remind us of its contextually limited explanatory power; in complex systems, “who or what actually perceives the meaning of situations so as to take appropriate actions? This is essentially the question of what constitutes consciousness or self-awareness in living systems … [This is one] among the most profound mysteries in complex systems and in science in general. Although this mystery has been the subject of many books of science and philosophy, it has not yet been completely explained to anyone’s satisfaction” ([
58], p. 184). Hence, the objections to MM raised here also impinge on scientific attempts at using the FEP to understand the emergence of inference, representation or meaning. More specifically, the FEP is explanatory as a way of making predictions “as if” an extant physical system behaves with a dual geometry. However, the FEP does not explain the existence of inference, representation or meaning in nature.
8. Conclusions
Throughout the previous sections, this paper has submitted the FEP framework (
Section 2) to profound philosophical inspection. It has nuanced previous examination of the FEP from Colombo and Wright [
12] in the following sense: whereas general criticisms of the FEP as a GUT in terms of epistemic pluralism seem misplaced, since they tend to cut off the spirit of scientific reduction and theoretical scientific progress (
Section 3), other criticisms from mechanists and organicists stand and ask for deeper metaphysical scrutiny (
Section 4). The opportunity for such scrutiny comes on the back of FEP theorists embracing philosophy in a recent publication [
13], where they defend so-called Markovian Monism (MM) (
Section 5). This is good news for interdisciplinary work between science and philosophy without explicitly requiring scientific explanatory pluralism.
MM was presented and reviewed in this paper by making explicit the FEP’s implicit assumptions, which Friston and colleagues insufficiently disclosed (
Section 6). In particular, MM does not tackle the emergence of representation through evolution, taking for granted some living systems’ specific cognitive ability. More importantly, MM veils the assumptions behind separating system and environment via a Markov blanket. Such drawbacks heavily stymie a FEP interpretation consistent with MM. Nonetheless, inasmuch as the FEP itself is not committed to MM, such critiques do not invalidate the epistemic advancement that the FEP framework provides. Even though the FEP may incur in circular reasoning, this drawback simply weighs upon its Markovian monist interpretation, and not necessarily on its value as a generalized heuristic principle for better characterization of homeostatic phases in already extant biological systems (
Section 7). The FEP’s recourse to Markov blankets induces circular causalities between living systems and the environment, which, freed from MM interpretations, can go beyond circular reasoning.