Feature Extraction Using an Unsupervised Neural Network

摘要

A novel unsupervised neural network for dimensionality reduction which seeksdirections emphasizing distinguishing features in the data is presented. A statistical framework for the parameter estimation problem associated with this neural network is given and its connection to exploratory projection pursuit methods is established. The network is shown to minimize a loss function (projection index) over a set of parameters, yielding an optimal decision rule under some norm. A specific projection index that favors directions possessing multimodality is presented. This leads to a similar form to the synaptic modification equations governing learning in Bienenstock, Cooper, and Munro (BCM) neurons (1982). The importance of a dimensionality reduction principle based solely on distinguishing features, is demonstrated using a linguistically motivated phoneme recognition experiment, and compared with feature extraction using principal components and back propagation network.