Rosa Zimmermann: Self-Timed Self-Supervised Learning in Networks of Tempotrons

BCCN Berlin / Technische Universität Berlin

Abstract

Self-supervised networks composed of leaky-integrate-and-fire units can detect recurring features in spiking input patterns when using the Tempotron learning rule and computing the training label as a measure of central tendency of the output spike count of the member units. The published version of the self-supervised networks of Tempotrons relies on the availability of a set of individual trials. In this thesis such networks are extended to continuous input streams in order to make the model applicable to a wider range of scenarios. To this end, a Tempotron-network is introduced which can decide at which points in the input stream the member units of the network should perform learning steps without requiring a way to measure the objective passage of time, i.e. a self-timed network is created. The notion of a continuum of supervisory behaviour that covers the range between the original and the self-timed network is developed. Finally, a stochastic description of the networks in a fixed state is developed and used to evaluate a measure of the quality of supervision of a given network variety across contiguous sub-spaces of network state space. The resulting landscapes of supervision quality are examined and compared for multiple different ways to calculate the target towards which the members of the network should adjust their response. It is demonstrated that the use of the mean, rather than the mode, renders the network more susceptible to being equidistant to two solutions and causes great stochasticity in the trajectories networks take through state space.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Robert Gütig   & Prof. Tilo Schwalger    / Lisa Velenosi