Stefan Rotter: Associative rewiring and repair in self-organizing networks

Abstract

Configurable neural networks in the brain are generally thought to provide a robust and flexible basis for memory and learning. This is equally true for invertebrates and vertebrates, including mammals and humans. In recent decades, a variety of rules for synaptic plasticity and related network-wide mechanisms have been proposed to explain experimental observations. However, we are still far from being able to explain psychological learning phenomena with a comprehensive network-based theory.

Our first goal was to develop a mathematical theory of graphs that dynamically evolve over time and respond to perturbations in their structure. Within this framework, the question then arose as to what additional assumptions would provide a plausible minimal theory of learning in biological networks as they interact with the outside world. Two-photon imaging of structural synaptic plasticity and, more recently, observations of molecular processes during functional plasticity at the nanoscale provide the empirical basis and further inspiration for this endeavor. Our approach has been to describe the presumed subcellular processes and to derive macroscopic predictions from this description.

The basic idea is that synaptic networks self-organize entropically, following the law of mass action known from chemistry. This approach has indeed led to a useful description, the dynamic directed configuration model. I will describe some emergent properties of this new model, including Hebb's rule (“cells that fire together wire together”) and self- repair of networks under appropriate conditions. I will also discuss possible links to psychological phenomena and describe new gradient-free machine learning strategies that can be derived from our biologically motivated theory.

About the speaker and his research

Hosted by Christian Leibold