Colloquium Cognitive Systems

Title: Bounded rationality and beyond: information-theoretic models of learning and decision-making in humans and machines

Prof. Dr. Dr. Daniel Braun (Ulm Universtiy)

Abstract:
Decision theory tells an agent what to do once its objective and beliefs are given. It does not consider the cost of computing the best answer for a given objective, nor what to do when no objective is given at all. In the talk we address the first question through information-theoretic bounded rationality, which treats information processing as a resource and trades off expected utility against the cost of reducing uncertainty. We will discuss some of the evidence showing how such a trade-off can capture the efficiency of human reaching movements, the emergence of abstractions and heuristics, neural signatures of information costs, and deviations from Nash equilibria in sensorimotor interactions. For the second question we will discuss recent work on agents that construct their knowledge in the absence of a given objective, for example by learning goals as partial orders, by learning from intended versus reached goals, and by attending only to what remains predictable. Finally, we ask to what extent information itself depends on an objective. We conceptualize knowledge as a partition of the possible states of the world, where partitions form a lattice that exists before any objective is given, while information is the risk reduction between two knowledge states once an objective is fixed. In this picture, bounded rationality describes how far an agent should move up the lattice when each step costs a resource, while the formation of knowledge itself requires new methods.
  
About:
Daniel Alexander Braun studied physics (Dipl. 2005), biology (Dipl. 2005) and philosophy (Cand. phil. 2005), and received doctorates in natural science (Dr rer nat 2008) and philosophy (Dr phil 2011) at the Albert-Ludwigs-Universität, Freiburg, in the subject areas of computational neuroscience and philosophy of mind respectively. He received his habilitation in neural and behavioral biology (2014) and cognitive science (2015) at the Eberhard Karls Universität Tübingen. Previously, he was a visiting PhD student and postdoctoral research associate in the Computational and Biological Learning Laboratory (2006-2010) at Cambridge University and a visiting scientist (2010-2011) in the Computational Learning and Motor Control Laboratory at the University of Southern California. In 2011 he was awarded an Emmy-Noether-fellowship by the Deutsche Forschungsgemeinschaft to establish the independent research group “Sensorimotor Learning and Decision-making” at the Max-Planck-Institutes for Biological Cybernetics and Intelligent Systems in Tübingen. In 2015 he was awarded an ERC Starting Grant “BRISC: Bounded rationality in sensorimotor coordination”. Since 2016 he is Professor of Learning Systems at Ulm University.