Active inference: building a new bridge between control theory and embodied cognitive science

The application of Bayesian techniques to the study and computational modelling of biological systems is one of the most remarkable advances in the natural and cognitive sciences over the last 50 years. More recently, it has been proposed that Bayesian frameworks are not only useful for building descriptive models of biological functions, but that living systems themselves can be seen as Bayesian (inference) machines. On this view, the statistical tools more traditionally used to account for data in biology, neuroscience and psychology, are now used to model the mechanisms underlying functions and properties of living systems as if the systems themselves were the ones“calculating”those probabilities following Bayesian inference schemes. The free energy principle (FEP) is a framework proposed in light of this paradigm shift, advocating the minimisation of variational free energy, a proxy for sensory surprisal, as a general computational principle for biological systems. More intuitively and under some simplifying assumptions,the minimisation of variational free energy reduces,for an agent,to the minimisation of prediction errors on sensory input. Initially proposed as a candidate unifying theory of brain functioning, the FEP was later extended to encompass hypotheses on the origins of life, and is nowadays discussed in the cognitive science community for its possible implications for theories of the mind. In particular,one of the most popular process theories derived from the FEP,active inference,describes a biologically plausible algorithmic implementation of this principle with several repercussions on our understanding of cognition. In this thesis, I will focus on the role of this process theory for action and perception. In active inference, the two of them are combined in a closed sensorimotor loopasco-dependent processes of minimisation of a single loss function,variational free energy, with respect to different sets of variables. Building on this, I will suggest that some of the core ideas of active inference are best seen in terms of enactive, embodied, extended and embedded (4E) theories, in contrast to the majority of the literature emphasising its apparent connections to more traditional, computational, accounts of the mind. In particular, I will develop this argument by focusing on some proposals central to 4E approaches: (a) the non-brain-centric nature of cognitive processes,(b)the lack of explicit representations of the world,(c)the coupling of agent-environment systems and (d) the necessity of real-time feedback signals from the environment. Under the FEP formulation, I will present a series of case studies with mainly two objectives in mind: 1) to conceptually analyse and reframe these 4E ideas in the context of active inference, arguing for the advantages of their formalisation in a more general probabilistic (Bayesian) framework and, 2) to present new mathematical models and agent-based implementations of some of the conceptual connections between Bayesian inference frameworks and 4E proposals, largely missing in the literature.