As mentioned above, all this serves the basic function of the brain: the minimization of free energy. This strategy is employed in every aspect of cognition; thus the free energy principle (Friston 2010) is a grand unifying theory. But from where does the free energy principle derive its unifying power?[2]
The free energy principle makes use of Bayesian theory, which can be regarded as its foundation. For some years now, Bayesian theory has been applied to many cognitive phenomena, since it may “offer a new interpretation of the goals of cognitive systems, in terms of inductive probabilistic inference […][,] can make the assumptions of Bayesian models more transparent than in mechanistically oriented models […][and] may have the potential to explain some of the most complex aspects of human cognition […]” (Jones & Love 2011, p. 170). Yet Jones & Love (2011) also address the fact that Bayesian theories, although aiming at researching and investigating the human brain and its workings, remain unconstrained by psychology and neuroscience “and are generally not grounded in empirical measurement” (ibid., p. 169). They term this approach “Bayesian Fundamentalism”, since it entails that all that is necessary to explain human behaviour is rational analysis. Supporters of this position rely on the mathematical framework of Bayesian theory as the origin of its explanatory power and unification. The positive thesis of Jones & Love (2011) consists in arguing for “Bayesian Enlightenment” that tries to include mechanistic explanation in Bayesian theory. To give more detail, they propose that, rather than following Bayesian Fundamentalism and thus being “logically unable to account for mechanistic constraints on behavior […] one could treat various elements of Bayesian models as psychological assumptions subject to empirical test” (Jones & Love 2011, p. 184). Similarly, Colombo & Hartmann (2014) argue that although “the Bayesian framework […] does not necessarily reveal aspects of a mechanism[,] Bayesian unification […] can place fruitful constraints on causal-mechanical explanation” (Colombo & Hartmann 2014, p. 1).
According to Colombo & Hartmann (2014), many Bayesian theorists falsely equate unification with explanatory power. But Bayesian theories derive their unificatory power from their mathematical framework. However, just because different cognitive phenomena can be mathematically unified does not entail a causal relationship between them, and nor does the mathematical unification tell us anything about the causal history of these phenomena. However, as will be presented in the next section, explanatory power, at least from a mechanistic point of view, results from investigating structural components and their causal interactions that give rise to a certain phenomenon. For example Kaplan & Craver (2011) write that “[…] the line that demarcates explanations from merely empirically adequate models seems to correspond to whether the model describes the relevant causal structures that produce, underlie, or maintain the explanandum phenomenon” (p. 602). Yet in the case of Bayesian theory—and Bayesian Fundamentalism in particular—, this cannot be achieved, since they “say nothing about the spatio-temporally organized components and causal activities that may produce particular cognitive phenomena […]” (Colombo & Hartmann 2014, p. 5). But not everything is lost concerning the explanatory role of Bayesian theories. Even if Bayesian theory cannot provide mechanistic explanations, it may nonetheless be beneficial to cognitive science by offering constraints on causal-mechanical explanation (Colombo & Hartmann 2014).
This brings us to the free energy principle. As noted, the free energy principle is, at its core, a theory that makes use of Bayesian theory; consequently it inherits all of Bayesian theory’s pros and cons. Thus, since unification in the free energy principle is also grounded in its mathematical foundations “[…] the real challenge is to understand how [the free energy principle] manifests in the brain” (Friston 2010, p. 10). With regard to Jones & Love’s (2011) distinction, the free energy principle can be considered to belong to Bayesian Enlightenment, since it attempts to ground its findings in neurobiology and psychology rather than remaining unconstrained by these sciences. Furthermore, due to the fact that the free energy principle integrates neuroscientific findings into its conclusions, it can offer more precise constraints on causal-mechanical explanations than Bayesian theory alone. For example, the free energy principle tries to incorporate neuroscientific facts about brain structure and its hierarchical organization, or tries to link concepts such as “precision” to neurophysiological phenomena such as “dopaminergic gating” (Friston et al. 2012).[3] The latter example will be presented in greater detail in section 5.
In sum, the free energy principle offers a form of unification that exceeds that offered by Bayesian theory alone. It makes statements about how the free energy principle could be realised in the brain and does not solely rely on its mathematical framework. Thus, one could term the former a “strong unification thesis” (SUT) and the latter a “weak unification thesis” (WUT).
If the free energy principle is true it creates a backdrop against which other theories must be evaluated. This also implies a kind of explanatory monopolization, since “the free energy principle is not a theory that lends itself particularly well to piecemeal” (Hohwy this collection, p. 9). In other words, as Hohwy highlights on many occasions, the free energy principle is an all-or-nothing theory. He compares it to the theory of evolution in biology and states that, just like the free energy principle, “evolution posits such a fundamental mechanism that anything short of universal quantification would invalidate it” (p. 10). Due to this large explanatory ambition, some researchers have described the free energy principle as preposterous. Yet “the issue whether the free energy principle is preposterous cannot be decided just by pointing to its explanatory ambition […] [but] by considering the evidence in favour of the free energy principle” (p. 11). This is a very important transition, i.e., the switch from explanatory ambition to explanatory power, since, from a mechanistic viewpoint, the former gives no statement about the veridicality of its assumptions, whereas the latter does.
In the remainder of this paper, I will argue that one major shortcoming of the free energy principle lies in its explanatory power. The main issue to be discussed consists in the fact that most concepts employed in the free energy principle, or in its applications such as predictive coding (Friston 2005; Rao & Ballard 1999) or predictive processing (Clark 2013; Hohwy 2013), are principally functional concepts. Yet, at least in the case of the free energy principle, functional concepts do not hold much explanatory power, since they “describe how things ought to work rather than how they in fact work” (Craver 2013, p. 18). For example, the concept of “precision” represents the amount of uncertainty in the incoming sensory signal that may arise due to noise. Thus the precision of the incoming sensory inputs determines how an agent interacts with its environment next: it can either change its models or its sensory input. Yet, this description holds no commitments as to how precision is realised in the brain; it only describes what effect precision should have on a given cognitive system. Therefore the free energy principle seems to be of a normative, rather than descriptive, nature.[4] On the other hand, there are mechanistic explanations that, according to Craver (2007), can also count as such, since they don’t describe how things should work but how they in fact do work.
Yet these two types of epistemic strategies don’t necessarily exclude each other. Here I want to introduce Piccinini & Craver’s (2011) claim that functional analyses can serve as “mechanism sketches”. The upshot lies within the free energy-principle’s unifying power: it can act as a kind of conceptual guide for revealing mechanistic explanations. Once physiological concepts are mapped onto the functional concepts derived from the free energy-principle, multilevel mechanistic explanations follow. But before this is elaborated the next section will give a short introduction to mechanistic explanation (Craver 2007).