Max Planck Institute for Dynamics and Self-Organization -- Department for Nonlinear Dynamics and Network Dynamics Group
Personal tools
Log in


Monday, 12.01.2015 14 c.t.
Date changed!

Understanding information processing goals in neural systems with partial information decomposition

by Prof. Dr. Michael Wibral
from Brain Imaging Center, MEG Unit, Frankfurt

Contact person: Viola Priesemann


Ludwig Prandtl lecture hall


Different tasks in neural systems seem to be solved by similar anatomical motifs, with six layered mammalian cortex as a prime example of such a motif. This observation spawned interest in common principles of information processing implemented by these motifs. These common principles by definition cannot be cast in processing-domain specific language (e.g. detection of domain specific features). Thus, our aim is to formulate and verify the principles of neural information processing in terms of information theory. This task can be simplified using the recently introduced concepts of partial information decomposition (PID). PID separates unique, redundant and synergistic information contributed by the inputs to a neural processor to its output. The talk will introduce the main ideas of PID and then present a generic information theoretic goal function that can represent various popular processing strategies, such as predictive coding, infomax and the coherent infomax principle, by a simple choice of parameters. Last we show how to map some of these goal functions to the neural implementation of coherent infomax to reuse its well studied activation functions and learning rules.


Date changed!

back to overview