Title :
A neural network based on LVQ2 with dynamic building of the map
Author :
Maillard, E. ; Solaiman, B.
Author_Institution :
TROP Lab., Univ. de Haute Alsace, Mulhouse, France
fDate :
27 Jun-2 Jul 1994
Abstract :
HLVQ network achieves a synthesis of supervised and unsupervised learning. Promising results have been reported elsewhere. A dynamic map-building technique for HLVQ is introduced, During learning, the creation of neurons follows a loose KD-tree algorithm. A criterion for the detection of the network weakness to match the topology of the training set is presented. This information is localized in the input space. When the weakness criterion is matched, a neuron is added to the existing map in a way that preserves the topology of the network. This new algorithm sets the network almost free of a crucial external parameter: the size of the neuron map. Furthermore, it is shown that the network presents highest classification score when employing constant learning rate and neighborhood size
Keywords :
learning (artificial intelligence); neural nets; vector quantisation; HLVQ network; LVQ2; constant learning rate; constant neighborhood size; dynamic map-building; loose KD-tree algorithm; neural network; supervised learning; training set topology; unsupervised learning; Convergence; Data compression; Laboratories; Network synthesis; Neural networks; Neurons; Pattern classification; Telecommunication network topology; Unsupervised learning; Vector quantization;
Conference_Titel :
Neural Networks, 1994. IEEE World Congress on Computational Intelligence., 1994 IEEE International Conference on
Conference_Location :
Orlando, FL
Print_ISBN :
0-7803-1901-X
DOI :
10.1109/ICNN.1994.374274
