Unsupervised Learning and Self-Organization in Networks of Spiking Neurons

Abstract

One of the most prominent features of biological neural systems is that individual neurons communicate via short electrical pulses, the so-called action potentials or spikes. In this chapter we investigate possible mechanisms of unsupervised learning and self-organization in networks of spiking neurons. After giving a brief introduction to spiking neuron networks we describe a biologically plausible algorithm for these networks to find clusters in a high dimensional input space or a subspace of it. The algorithm is shown to work even in a dynamically changing environment. Futhermore, we study self-organizing maps of spiking neurons showing that networks of spiking neurons using temporal coding can achieve a topology preserving behavior quite similar to that of Kohonen's self-organizing map. For these networks a mechanism of competitive computation is proposed that is based on action potential timing. Thus, the winner in a population of competing neurons can be determined locally and in generally faster than in approaches which use rate coding. The models and algorithms presented in this chapter establish further steps toward more realistic descriptions of unsupervised learning in biological neural systems.