Causal learning in the Crib: A predictive processing formalization and babybot simulation

Predictive Processing (PP) [1], [2], [3] is becoming an influential account in cognitive neuroscience, including developmental neuroscience [4]. According to PP, human brains interpret their sensory inputs by predicting them, based on a hierarchy of generative models. These predictions are then compared to the actual, observed inputs, and the difference between predictions and observations (so-called prediction error) is used to update the agent's generative model about the world, to minimize future prediction errors. A key question for PP is how situated agents can learn these generative models. This question is especially important from a developmental perspective on PP. That is, the theory needs to specify how such generative models can be `developable' at all, given that infants must somehow build these generative models from embodied, situated interaction with their environments.