Conference paper

Artificial Neural Networks - ICANN 2008

Part II

Springer

Berlin

2008

978-3-540-87558-1

Ing. Jiří Grim, CSc. (ÚTIA - Oddělení rozpoznávání obrazů)

Probabilistic neural networks, Feature extraction, Recogniti

In order to design probabilistic neural networks in the framework of pattern recognition we estimate class-conditional probability distributions in the form of finite mixtures of product components. As the mixture components correspond to neurons we specify the properties of neurons in terms of component parameters. The probabilistic features defined by neuron outputs can be used to transform the classification problem without information loss and, simultaneously, the Shannon entropy of the feature space is minimized. We show that, instead of dimensionality reduction, the decision problem can be simplified by using binary approximation of the probabilistic features. In experiments the resulting binary features improve recognition accuracy but also they are nearly independent - in accordance with the minimum entropy property.