Elisabeth Kress: Optimizing spiking neural networks for delayed-decision making tasks

BCCN Berlin / TU Berlin

 

Abstract

The capability to store and manipulate information within the activity of a neural population underlies the process of delayed-decision making. Modeling approaches to understand how this information is maintained over time in the vibrotactile discrimination task, have spanned methods with highly engineered dynamics, to loosely trained random networks. However, they have all been unable to resemble experimental characteristics of the active prefrontal cortex population and individual neuron tunings. All of these models rely on firing rate dynamics which ignores the temporal aspect of utilizing specific spike timing of neurons to transmit this stored information. We propose that by implementing spiking dynamics for a biologically closer network model, we find new hypotheses on computational techniques for working memory. First, comparing population characteristics in experimental recordings, rate, and spiking models, similarities and differences among the networks help evaluate model choices. Novel neuron tuning types are found in the rate model. In spiking networks, the population tuning to the stored information is most representative of the experimental population. Lastly, spike trains and initial hypotheses on spike timing dependencies are analyzed to contribute to early understanding of how spiking dynamics are utilized in storing information during working memory.

 

Additional Information

Master Thesis Defense

 

Organized by

Dr. Prof. Sprekeler and Dr. Prof. Obermayer