I’ll give an overview of my work on active inference and ideas derived from 4E (enactive, embodied, embedded and extended) cognition in the last few years, highlighting what I think these theories can be used for and what, in my opinion, they cannot do. From (briefly) discussing the standard idea of generative models inherited from statistics and machine learning to deconstructing the idea of “generative” in generative models for biology and cognitive science. From the mathematical overlaps of theories of inference and control to an example application of active inference as PID control in control theory/engineering (maybe biology and cognitive science too?). From the idea of “doing philosophy of mind with a screwdriver”, to some attempts of creating a common language for discussions in cognitive science based on predictive processing and active inference.

Where does our ability to make abstraction arise? Humans excel in creating concepts, organizing information in schemas and hierarchies, which in turn allow efficient learning and behaviour to emerge. We hypothesized that value assigned to sensory features is key in driving abstraction to subserve learning. We utilize two computational models differing in their state-space dimensionality to characterize participants’ behavioral strategies in learning the fruit preference of imaginary characters. We show that participants with higher abstraction over the task also solve the task faster, validating the recurring argument that abstraction means efficiency. Participants’ learning ability is predicted by the functional connectivity between visual areas and ventromedial prefrontal cortex (vmPFC), where value is thought to be computed. Moreover, we find that feature-level information is represented in the hippocampus and vmPFC when behavior is ‘abstract’ – confirming the intuition that abstraction may result from the combination of sensory information with a ‘goal-dependent’ value signal. Finally, by reinforcing the neural occurrence of a target feature in visual cortex we show that this leads to increased behavioral abstraction. I will discuss these results within the broader context of current theories of value computation, abstraction and reasoning.