Fabry provides a persuasive case for how PP could provide the neural underpinnings of enculturation. In this section, however, I will raise some problems for the proposed marriage of CI/ENC and PP. The main issues I will address are as follows:
The incompatibility of the isolated brain interpretation (Hohwy 2013) and the active inference interpretation (Clark 2013) of PP.
The attempt to explain all cognitive processing in terms of prediction error.
The redeployment of neural circuitry as not being explained by PP.
The role of NPP as not being explained by the reduction of prediction error.
1. Isolating the brain
If CI/ENC has one central commitment, it is that we should not think of cognition as isolated from the environment. And yet this is exactly how we ought to understand the predictive brain, according to a prominent interpretation of the PP framework. Whenever the PP framework is introduced, it is almost always introduced in the following way: “Accounts of PP generally assume that human perception, action, and cognition are realized by Bayesian probabilistic generative models implemented in the human brain. Since the human brain does not have immediate access to the environmental causes of sensory effects, it has to infer the most probable state of affairs in the environment giving rise to sensory data” (Fabry 2015, p. 4; my emphasis). The two main motivations for the PP framework are that the brain is isolated from the environment and must make a best guess as to what it is perceiving, and that this kind of probabilistic inference-making results in internal (neurally realized) models of the environment. Putting aside the probabilistic nature of the inferences, this just is old-fashioned individualism. There is a perceptual interface to an environment of hidden variables; the internal system creates internal models (representations) of those hidden environmental variables, which then causally produce behaviour. The internal states must predict the external variables via sensory input, but they have no direct access to the causal ancestry of the sensory input. This form of individualism is used as an explanation for why models and predictions are required: “Because the brain is isolated behind the veil of sensory input, it is then advantageous for it to devise ways of optimizing the information channel from the world to the senses” (Hohwy 2013, p. 238). Hohwy describes the mind–world relation as “fragile” because of the isolation of the brain, and this is why active inference is required.
The saving grace of the PP framework, from the perspective of CI/ENC, is active inference. In Clark’s version of PP active inference and cultural props help to minimize prediction errors (Clark 2013); and because of this, there is a deep continuity between mind and world mediated by active inference and the cultural scaffolding of our local niche. Curiously, Hohwy agrees with Clark’s interpretation, but at a cost. Hohwy agrees that active inference and the cultural scaffolding of the environment help to change sensory input so as to minimize prediction error, but also “by increasing the precision of the sensory input” (Hohwy 2013, p. 238). According to Hohwy, the primary role of PP is perceptual inference; as a matter of “second order statistics” active inference helps to optimise sensory input so that perceptual inference is less error-prone.
Note the cost. First, active inference and cultural scaffolding is relegated to the secondary role of reducing prediction error for the primary cognitive job of perceptual inference, which is carried out wholly by matching statistical models to sensory input in the brain. Second, Hohwy shows that this interpretation of active inference should be understood against the background of the isolated brain. “The key point I am aiming at here is that this is a picture that accentuates the indirect, skull-bound nature of the prediction error minimization mechanism” (Hohwy 2013, p. 238). Organizing and structuring our environments makes sense if the mind–world relation is fragile in the way that Hohwy presents it, and also because this structuring makes perceptual inference more reliable. I take it that Fabry and Clark would deny this interpretation of the role of active inference and cultural scaffolding. Indeed, Fabry denies Hohwy’s ‘isolationist’ interpretation in her commentary.
However, Fabry does so by playing up the roles of NPP, which go far beyond prediction minimization: “Furthermore, we need to take into account that genuinely human cognitive processes occur in a culturally sculpted cognitive niche. […] These cognitive resources have unique properties that render them particularly useful for the completion of cognitive tasks” (Fabry 2015, p. 12). She also nods to the sub-personal, mechanistic role of PP in the entire brain–body–niche nexus: “[T]he important theoretical contribution made by the prediction error minimization framework is its providing of a sub-personal, mechanistic description of the underlying neuronal and bodily sub-processes” (Fabry 2015, p. 13). It is therefore not clear to me that PP does anything more than provide the functional details of some of the neural processing in the brain–body–niche nexus. It certainly should not be taken to provide a comprehensive account of what cognition is and why there is cultural scaffolding, or what its interesting cognitive properties are.[5] It is to these issues that I shall now turn.
2. Everything is predicted
One of the main concerns with the PP approach is that it is used both to try to explain all of cognition and as an explanation of why there is cultural scaffolding. We’ve already seen a brief hint of this in Hohwy, Clark, and Fabry’s work above.[6] The first worry can be found in the expression of PP as originating in the free energy principle:
The free-energy considered here represents a bound on the surprise inherent in any exchange with the environment, under expectations encoded by its state or configuration. A system can minimise free energy by changing its configuration to change the way it samples the environment, or to change its expectations. These changes correspond to action and perception, respectively, and lead to an adaptive exchange with the environment that is characteristic of biological systems. This treatment implies that the system’s state and structure encode an implicit and probabilistic model of the environment. (Friston & Stephan 2007, p. 417)
PP is primarily a model of the way in which top-down processing ‘predicts’ bottom up sensory input and which samples the environment to change its expectations. These correspond to perception and action respectively[7]. However, it seems odd to build a cognitive theory on the basis of the prediction of sensory signals. This is because much of cognition is not about sensory signal prediction; nor about actions as sampling the environment. Indeed much of cognition isn’t about ‘prediction’ at all. So whilst I agree that at least part of the mechanisms of cognition can be fruitfully modelled by PP, not all of them will be. In enculturated systems, the really important work is being done by the processing governed by normative patterned practices whose properties are understood primarily at the social or populational level. I agree that at the individual level, the mechanisms of ICS can partly be explained by PP, but the main explanatory work will not be a matter of predictions of sensory input[8].
The examples from Landy & Goldstone (2007) may be partly explained by prediction errors, but again this only makes sense in the context of sensorimotor processing governed by mathematical norms. If the norms function as priors in the system, then this might help explain the errors made by the test subjects.
3. Phenotypic plasticity and neural redeployment
PP can’t explain the redeployment of neural circuitry to new cognitive functions. And it is not supposed to, since this isn’t the job it was designed to do. However, this is a considerable weakness if PP is supposed to be the primary mechanism of enculturation. I've already canvassed the reasons why in section 1.
4. NPP and prediction error minimization
Enculturated PP plays a role in the multi-layered and interwoven ICS, but it neither determines nor implements the entire system. My argument in this response has been that the dynamics of ICS are not determined by the predictive processing of parts of the system: if anything PP is enslaved to the processing needs of the entire enculturated system. The PP framework takes perceptual inference as its primary mode of processing, which is the top-down matching of predictions to sensory input. However, it is not obvious that this is the right model for all cognitive processing, since it is not obvious that all cognitive processing is just a matter of predictions about sensory input, nor a hierarchically organised system which minimises prediction errors.
For example Hohwy (2013, p. 238) argues that “many of the ways we interact with the world in technical and cultural aspects can be characterized by attempts to make the link between the sensory input and the causes more precise (or less uncertain).” This would be a very impoverished account of the evolution of public systems of representation. Public systems of representation did not simply evolve to “make the link between the sensory input and the causes more precise (or less uncertain)”; this would be to ignore the social pressures that would have caused representational innovation.[9] It might be true that the history of the refinement of notation has something to do with making input more easy to ‘predict’; however, this would not be an ultimate explanation for why there are notations in the first place, nor how they function in our cognitive lives. It might be a proximal explanation of the neural mechanisms for the processing of notations and as such, it might explain some of the causal conditions that explain how notations have developed, but it doesn’t explain the conditions under which notations evolved. For further reasons why see section 3.4 of my target article, on evolutionary novelty and uniqueness (this volume).
For example, the idea that the brain predicts the product of two numerals makes sense, and the surprise at a product too distant from the operands lends further credence. Remember the example from section 4.1 of my target article (this volume) : 34 + 47 = 268. However, it is not obvious that predictions will help with the second example: 34 x 47 = 1598. What is required in this instance is the serial working through of the multiplication according to an algorithm. Furthermore, this is not simply a case of sensory predictions: when it comes to recognising the numerals on the page in front of you, PP can explain top-down predictions about sensory input, but that is not at all the same thing as the working through of a mathematical problem. So mathematical cognition could not, it seems to me, be reducible to error minimization.