Critchley, David A.; (1994) Extending the Kohonen Self-Organising Map by Use of Adaptive Parameters and Temporal Neurons. Doctoral thesis (Ph.D), UCL (University College London).

Text Extending_the_Kohonen_self-org.pdf Download (10MB)

Abstract

This work extends the Kohonen self-organising map in two primary ways: o A dynamic extension to the model which allows the neighbourhood size and learning rate timecourse to be deduced during learning. o Inclusion of temporal features, both in single layer and hierarchical networks. The dynamic learning parameter model is developed as a consequence of how the self-organising map forms 'stable states' under fixed values of the learning parameters whilst exposed to a driving probability distribution. Such stable states can be used to deduce an appropriate stage to make a transition to a new set of learning parameters. This leads to a sequence of states that ultimately result in convergence. Temporal features are developed in the light of the Temporal Kohonen Map model of Chappell and Taylor. It is shown that application of the standard Kohonen learning law to such a network can lead to instability in the weightspace. This problem is shown to be soluble by moving the integrating characteristics from the cell body (where it is a scalar quantity) to the synapses (where it is a vector quantity). Multilayer temporal topographic mappings are discussed in terms of coding strategies between layers. The codings examined include complete feed-forward, feed-forward with enforced output spectrum and 'triangular coding', a binary coding of topology.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item