**Nihat Ay**

Max Planck Institute for Mathematics in the Sciences Leipzig, Germany

**Information-geometric structures in cognitive systems research**

**Watch video**

**Download video**

**Abstact:**

I will introduce the sensorimotor loop of an agent in terms of a causal model and highlight corresponding information-geometric structures. It turns out that the geometry of policy models can have a crucial effect on the quality of learning processes. I will discuss more explicitly the geometry of conditional restricted Boltzmann machines, in particular its relation to the universal approximation property. I will contrast this property with the notion of ?cheap control? within the field of embodied intelligence. This notion highlights the fact that high behavioural complexity, seen from the external observer perspective, does not necessarily imply high control complexity. I will present results on the geometric design of embodied systems with concise control architectures. My talk is based on a number of works in collaboration with Guido Montufar, Keyan Ghazi-Zahedi, and Johannes Rauh.