In this paper, two performances increasing methods for datasets which have a nonuniform class distribution are presented. The methods are applied to probabilistic neural networks (PNN). Selection of a good training data is the most important issue. Therefore, a new data selection procedure including data exchange and data replication is proposed. After reaching the best accuracy by using the data exchange method, a data replication method is applied to the classes which have relatively less numbers of instances. The methods are applied to the Glass, Escheria Coli (E. coli) and Contact Lenses datasets, which have nonuniform class distributions and better accuracies than the reference works were achieved by PNN using these methods.