090209 Morrison

'Learning' is an aspect of cognitive function that typically bridges multiple levels of description, with synaptic plasticity mechanisms at one extreme and organism behavior at the other. For this reason it is a difficult topic to address; top-down approaches are criticized because they are often formulated in such an abstract fashion that it is not clear how they could be implemented in a neural substrate, whereas bottom-up approaches generally fail to generate high level functionality.
Consequently, a crucial tool for understanding learning is the simulation of biological neuronal networks. This allows us to produce concrete realizations of learning theories, informed and constrained by analyzing the results of physiological experiments, thus connecting the top-down and bottom-up approaches. In this talk, I will present the results of two projects dealing with different facets of learning that demonstrate the usefulness of this combined approach.
The first project is a study of spike-timing dependent plasticity (STDP), in particular its influence on the stability of large-scale recurrent networks and on the generation of functional structure. Secondly, I will present a spiking neural network that implements temporal-difference learning on the basis of fluctations of a dopamine signal which is dynamically generated by the network itself. The size and complexity of the studied systems demands the use of high-performance computing and specially designed software. I will briefly outline the major features of the simulation tool we developed for this purpose.