Combining Multiple Classifiers in Probabilistic Neural Networks Jiˇ r´ ı Grim 1 , Josef Kittler 2 , Pavel Pudil 1 , and Petr Somol 1 1 Institute of Information Theory and Automation, P.O.BOX 18, CZ-18208 Prague 8, Czech Republic, grim@utia.cas.cz, pudil@utia.cas.cz, somol@utia.cas.cz 2 School of Electronic Engineering, Information Technology and Mathematics, University of Surrey, Guildford GU2 5XH, United Kingdom Abstract. We first summarize main features of a new probabilistic ap- proach to neural networks recently developed in a series of papers in the framework of statistical pattern recognition. We consider a simplifying binary approximation of the output variables and, in order to prevent the arising information loss, we propose to combine multiple solutions. How- ever, instead of combining different a posteriori probabilities, we make a parallel use of the binary output vectors to compute the standard Bayesian classifier. 1 Introduction The probabilistic approach to neural networks is closely related to statistical pat- tern recognition. The fundamental idea of probabilistic neural networks (PNN) is to approximate the class-conditional distributions by finite mixtures and to identify the components of mixtures with neurons (cf. e.g. [15, 16, 9, 6]). In the present paper we first summarize the basic principles of PNN. In order to pre- vent information loss in multilayer feed-forward PNN we consider a special type of classifier fusion. The standard approach to improve recognition accuracy and generalization performance of practical solutions, widely used both in statisti- cal pattern recognition [10, 11] and neural network ensembles [1, 7, 8, 12, 13] is to combine the outputs (e.g. a posteriori probabilities of classes, decisions) pro- duced by different classifiers. In the present paper we propose an alternative utilization of multiple classifiers which derives from the output representation adopted. Instead of applying various rules to different a posteriori probabilities of classes or output node excitations we make parallel use of multiple solutions by composing the corresponding binary subvectors. The parallel use of multiple solutions in the form of a joint binary output vector opens new possibilities to utilize the underlying decision information. The 0 Supported by the Grant of the Academy of Sciences No. A2075703, by the Grant of the Czech Ministry of Education No. VS 96063 and partially by the Complex research project of the Academy of Sciences No. K1075601 of the Czech Republic