Abstract
The free energy principle (FEP) has been presented as a unified brain theory, as a general principle for the self-organization of biological systems, and most recently as a principle for a theory of every thing. Additionally, active inference has been proposed as the process theory entailed by FEP that is able to model the full range of biological and cognitive events. In this paper, we challenge these two claims. We argue that FEP is not the general principle it is claimed to be, and that active inference is not the all-encompassing process theory it is purported to be either. The core aspects of our argumentation are that (i) FEP is just a way to generalize Bayesian inference to all domains by the use of a Markov blanket formalism, a generalization we call the Markov blanket trick; and that (ii) active inference presupposes successful perception and action instead of explaining them.
Highlights
Getting to the Free Energy Principle In the early 1990s, machine learning researchers developed a formalism analogous to Helmholtz free energy and began using it as an objective function for artificial neural networks in the context of parameter optimization through expectation-maximization algorithms (Dayan et al 1995; Hinton & van Camp 1993; Hinton & Zemel 1993; Neal & Hinton 1998; for early similar formulations, see Csiszár & Tusnády 1983; Hathaway 1986; Hinton & Sejnowski 1983)
The formalism was based on a quantity known as variational free energy with roots in the variational methods first developed by Richard Feynman (1972) and allowed researchers to solve otherwise intractable inferences when trying to guess the true distribution of partial sample data
As the free energy principle (FEP) story goes, once we understand the dynamics of systems in terms of their statistical properties and, more concretely, in terms of gradients over I(x), and we separate them from the external dynamics with a Markov blanket, we find a very interesting observation: the dynamics of the particular states of the system π can be seen as descending a gradient over I(x) defined in terms of external states η
Summary
Getting to the Free Energy Principle In the early 1990s, machine learning researchers developed a formalism analogous to Helmholtz free energy and began using it as an objective function for artificial neural networks in the context of parameter optimization through expectation-maximization algorithms (Dayan et al 1995; Hinton & van Camp 1993; Hinton & Zemel 1993; Neal & Hinton 1998; for early similar formulations, see Csiszár & Tusnády 1983; Hathaway 1986; Hinton & Sejnowski 1983). In the context of machine learning, algorithms based on the minimization of variational free energy and/or the equivalent ELBO maximization were used to optimize systems able to discover the latent/hidden variables (i.e., the true distribution) that generated the observed variables (i.e., the partial sample data)—e.g., what variational autoencoders do nowadays (Kingma & Welling 2014, 2019). In the case of cognitive systems, FEP speaks to their epistemological contact with their environment in terms of accurate perception and proper actions In all these cases, FEP entails a specific Bayesian process of minimization of free energy labelled as active inference. That active inference fails to account for perception and action; rather, it presupposes them Series of suggestions for issues that proponents of FEP and active inference will need to pay further attention to, especially if they wish to maintain the more ambitious claims they have made
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
