Configuring spiking neural network training algorithms

Mustari, Mst Mausumi Sabnam (2017) Configuring spiking neural network training algorithms. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (2MB)

Abstract

Spiking neural networks, based on biologically-plausible neurons with temporal information coding, are provably more powerful than widely used artificial neural networks based on sigmoid neurons (ANNs). However, training them is more challenging than training ANNs. Several methods have been proposed in the literature, each with its limitations: SpikeProp, NSEBP, ReSuMe, etc. And setting numerous parameters of spiking networks to obtain good accuracy has been largely ad hoc. In this work, we used automated algorithm configuration tools to determine optimal combinations of parameters for ANNs, artificial neural networks with components simulating glia cells (astrocytes), and for spiking neural networks with SpikeProp learning algorithm. This allowed us to achieve better accuracy on standard datasets (Iris and Wisconsin Breast Cancer), and showed that even after optimization augmenting an artificial neural network with glia results in improved performance. Guided by the experimental results, we have developed methods for determining values of several parameters of spiking neural networks, in particular weight and output ranges. These methods have been incorporated into a SpikeProp implementation.

Actions (login required)

View Item