Information Integration in Predictive Processes: A Mechanistic Grounding in the Self

The general aim of the project is to reveal necessary conditions for the emergence of internal representations associated with the self, when dealing with truly embodied and situated agents. This will be based on the study of Helmholtz machines that implement
prediction and recognition as prerequisite for optimal control. Furthermore, the aim is to study to what extent these processes generate high integrated information in the sense of Tononi’s Integrated Information Theory (IIT) of consciousness. This will provide insights about the mechanisms that underly the phenomenal self. Based on information theory, which is quantitative in nature, we expect to identify transitions between qualitatively  different kinds of embodiments, thereby relating our work to Metzinger’s orders of embodiment.
The work of this project will be based on crucial insights from the first period of the DFG SPP The Active Self which suggest refined research directions for the second period. In particular, we will develop a hierarchy of controller architectures with increasing granularity, ultimately leading to neuronal architectures. Here, we will benefit from our recent work on Helmholtz machines, which were originally proposed by Dayan et al. Corresponding
learning algorithms suggest a close connection to the Friston’s Free Energy Principle, which will provide a conceptual and formal basis for the project. Having developed controller architectures with various granularities, we will study corresponding information flows in sensorimotor loops of robotic systems, in collaboration with Verena V. Hafner’s group. We want to verify the increase of information integration within a controller when the learning
is involved and incorporates a forward model for prediction and an inverse model for control.