Models of Life and the Life of Models under the Free Energy Principle A cognitive realist perspective Majid D. Beni ([email protected]
) Introduction Science involves model-based reasoning, meaning that scientists use models of various stripes (mathematical, physical, phenomenological, etc.,) to represent the features of their target systems (Godfrey-Smith, 2006; Nersessian, 1999). In this context, one main issue is about realism--whether or not one can take a realist stance on models, or more importantly, whether scientific realism could be defended from within a model-based conception of science (Cartwright, 1999; Odenbaugh, 2011; Psillos, 2011). Very recently, the model-based approach to science has found its way to the centre of philosophical discussions of a flourishing theoretical framework in computational neuroscience. The theoretical framework is developed around the notion of the Free Energy Principle (FEP). This paper is not (only) concerned with realism about FEP models but (also) with how the understanding that FEP provides of mechanisms of the organism’s interaction with the world could provide fresh insights into scientific model making. FEP, as being developed by Karl Friston and colleagues (K. J. Friston, 2010, 2019; K. J. Friston et al., 2010; Ramstead et al., 2017) provide a vigorous theoretical framework to explain life and cognition in terms of the self-organising system’s tendency to remain in a limited number of places. A question is whether to take a realist or an instrumentalist tendency about FEP and its explanatory power (Colombo & Wright, 2016; Hohwy, 2014). More interestingly, in recent years the question of the realist fortitude of FEP (or lack thereof) has been entangled with the discussion of the model-based nature of FEP’s theoretical trajectory (Andrews, 2021; Beni, 2021a; Kirchhoff et al., 2022; van Es, 2020). This is where 1 we stand now: on one front, considerations of the model-based nature of the theoretical claims of FEP have been raised to substantiate an instrumentalist take on FEP (Colombo & Palacios, 2021; van Es, 2020; van Es & Hipolito, 2020), on the other, it has been remarked that despite the model-based nature of FEP, realism about FEP models is still a tenable option unless one wants to commit to what has been called a literalist fallacy (Kirchhoff et al., 2022). Well and good. But despite paving the way to realism about FEP is still a lively option, Kirchhoff et al.’s paper does not aim to offer a down-to-earth defence of realism about FEP models. The present paper does not offer a vigorous defence of realism about FEP either, but it aims to substantiate a version of cognitive realism about FEP by taking a convoluted path to realism. That is to say that the paper first offers an intelligible (ergodic) account of scientific model making by embracing the general rubric of FEP and then argues that since embracing FEP could underlie intelligible accounts of scientific reasoning, realism about FEP might be a lively option. As I say, although the present paper is still very much concerned with the relationship between the model-based nature of FEP and a realist philosophical stance on FEP, it takes a sophisticated path to explore the prospect of realism about FEP and its models; the paper starts its quest from the inspection of the nature of scientific models and asks how come that a tribe of human beings—called modellers or scientists—have the wonderful ability to model things—such as their biological and social environment as well as themselves. This means that, instead of directly engaging with the issue of realism about FEP, we start by focusing on the question of the origin of scientific models. Instead of asking “what is the relationship between the scientific models constructed using the FEP and the realities these models purport to represent?” (Kirchhoff et al., 2022, p. 2), we will ask what makes scientists’ attempt at making representational models of their environment so successful. I argue that not only the account of constructing generative models and minimising their conveyed prediction error provides a basis for explicating the origins of scientific model making it also helps with precisifying the notion of similarity in the context of model-based science. For, as I argue, the central notions of ‘similarity’ and ‘comparison’ in the context of model-based science (as explicated in the works of Ronald Giere et al.) have remained 2 vague and unspecified.