Here I will follow Piccinini & Craver’s (2011) proposal that functional descriptions are nothing other than mechanism sketches that derive their “[…] explanatory legitimacy from the idea that [they][…] capture something of the causal structure of the system” (Piccinini & Craver 2011, p. 306). Mechanism sketches are simply outlines of mechanisms that haven’t been fully investigated with regard to their structural properties. Thus, functional descriptions serve as placeholders until a mechanistic explanation can fully account for a given phenomenon by enriching functional concepts with concepts related to its structural properties.[5] The explanatory gaps[6] resulting from the functional nature of the free energy principle could then be closed, leading to a shift from explanatory ambition to explanatory power. This also directly relates to the alleged preposterousness of the free energy principle, since the process of “filling-in” will diminish any residual doubts about the theory’s truthfulness. This can be applied to the free energy principle, which works with functional concepts such as “precision”, “prediction error”, “model optimization” or “attention”: “[o]nce the missing aspects are filled in, a functional analysis turns into a full-blown mechanistic explanation” (Piccinini & Craver 2011, p. 284). Take the concept of precision in the free energy principle as an example. As described above, precision gives an estimate concerning the “trustworthiness” of a given sensory signal and its ensuing prediction errors. Taken as such, precision is clearly a functional concept since it is “[…] specified in terms of effects on some medium or component under certain conditions” (Piccinini & Craver 2011, p. 291) without committing to any structural entities that could realise these functional properties. However, according to Friston et al. (2012), “[…] dopaminergic gating may represent a Bayes-optimal encoding of precision that enhances the processing of particular sensory representations by selectively biasing bottom-up sensory information (prediction errors)” (p. 2). In turn, “dopaminergic gating” involves the neurotransmitter dopamine, a molecule that can be structurally described. Crucially, now that the functional concept of precision, derived from the free energy principle, has been linked with dopaminergic gating, one can make further inferences as to how this entity is situated in a multilevel mechanism. For example, the modulation of precision has been associated with attention (Feldman & Friston 2010; Hohwy 2012), and since precision is realised via dopamine mediation, one can investigate the effects of dopamine on attentional mechanisms.[7] On the other hand, if empirical evidence regarding precision or in particular predictions of precisions (hyperpriors) find “[…] that descending signals do not mediate expected precisions, this would falsify the free energy principle” (p. 16). This further accentuates the need for mechanistic explanations.
As a more elaborate example, the phenomenon of biased competition will shortly be introduced. In biased competition, two stimuli are presented at a topographically identical location. However, only one of these stimuli is actually perceived. Thus the principal question: by which means does the brain “select” any given stimuli? In the free energy principle, the most obvious answer would be the stimulus that best minimises free energy or prediction error. However, in these cases, the stimuli are equally accurate, i.e., they both represent the causal structure of the world equally well. As a consequence, the stimuli will “[…] compete for the responses of cells in visual cortex” (Desimone 1998, p. 1245). Crucially, Desimone (1998) brings up a preliminary study by Reynolds et al. (1994) that states “[…] that attention serves to modulate the suppressive interaction between two or more stimuli within the receptive field […]” (Desimone 1998, p. 1250). Thus, attention could be the determining factor as to which stimulus is perceived at a given moment. From the perspective of the free energy principle and in accordance with these findings, Feldman & Friston (2010) propose that “[…] attention is the process of optimizing synaptic gain to represent the precision of sensory information (prediction error) during hierarchical inference” (p. 2). These two views agree, since synaptic gain also entails a suppressive effect upon the other competing stimuli. Also, as just mentioned, Friston et al. (2012) identify precision weighting with dopaminergic gating, i.e., they argue that dopamine mediation realises the precision of incoming stimuli or prediction errors.
Now a fuller picture can be presented. This much more complete picture allows us to see how the free energy principle or prediction error minimization framework can prove to be beneficial with regard to mechanistic explanation. The phenomenon to be explained is biased competition. The mechanism that realises, or resolves, biased competition, i.e., the competition between two identically accurate and topographically identical stimuli, is precision weighting. This represents the etiological level of description since it describes how biased competition is resolved at a level of description that doesn’t refer to lower-level processes nor to how they are embedded into a higher order mechanism. It remains at the same level in the hierarchy of mechanisms. At the constitutive level we have the fact presented by Friston et al. (2012), that precision weighting is neurophysiologically realised by dopaminergic gating. This constitutes precision weighting and is located at a lower level. Last, precision weighting is embedded into the higher-order mechanism of attention. Precision weighting contributes to this higher order mechanism, or, from the other perspective, attention is constituted by precision weighting. This represents the contextual description.
The upshot is that, just as “[e]volutionary thinking can be heuristically useful as a guide to creative thinking about what an organism or organ is doing […]” (Craver 2013, p. 20), the free energy principle can be a useful guide in finding multilevel mechanistic explanations concerning how the mind works. Due to its unifying power, the free energy principle offers a grand framework that seeks to explain every aspect of human cognition. Thus, filling increasingly more mechanistic concepts into functional placeholders will enable an understanding of the mind in terms of how it does work instead of how it ought to work. The explanatory worth of the free energy principle would then be preserved, since “[i]f these heuristics contribute to revealing some relevant aspects of the mechanisms that produce phenomena of interest, then Bayesian unification has genuine explanatory traction” (Colombo & Hartmann 2014, p. 3).
However, this should not be seen as an attempt to eliminate functional concepts by reducing them to mechanistic ones. Instead, as mentioned above, the integrationist account emphasises that functional and mechanistic concepts are both necessary for mechanistic explanations, since “structural descriptions constrain the space of plausible functional descriptions, and functional descriptions are elliptical mechanistic descriptions” (Piccinini & Craver 2011, p. 307). Furthermore, once every functional term has a mechanistic counterpart, the 3M requirement posed by mechanists can be fulfilled in the case of the free energy principle.
Last, as a general remark, searching for structural properties seems important if researchers want to ground the free energy principle in the human brain. Functional theories are subject to multiple realizability. This means that not only humans or mammals could be bound to the free energy principle, but also Martians or bacteria or anything that could possess the “hardware” to do so. Hohwy suggests that the free energy principle can be seen as a biofunctionalist theory (this collection p. 20). In principle this means that the free energy principle can be multiply realised as long as that creature acts in such a way as to maintain itself in a certain set of expected states. These expected states then determine the creature’s phenotype. In seeking to explain human cognition, functional theories have to be enriched with mechanistic concepts relating to structural properties, since otherwise we could also be investigating Martians.