A NEW ONLINE LEARNING RULE FOR COMPETITIVE PRINCIPAL COMPONENTS ANALYSIS NEURAL NETWORKS

摘要

One of the best known techniques for multidimensional data analysis is the Principal Components Analysis (PCA). A number of local PCA neural models have been proposed to partition an input distribution into meaningful clusters. Each neuron of these models uses a certain number of basis vectors to represent the principal directions of a particular cluster. Most of these neural networks are unable to learn the number of basis vectors, which is specified a priori as a fixed parameter. This leads to poor adaptation to input data. Moreover, online learning is not supported in many of them. The PCA Competitive Learning (PCACL) is a method where the number of basis vectors of each neuron is learned online. Here we propose an improvement of its learning rule. Then we prove some important properties of the new rule. Finally, experimental results are presented where the original and modified versions of the neural model are compared.