This paper presents an approach to using both labelled and unlabelled data to train a multilayer perceptron. The unlabelled data are iteratively pre-processed by a perceptron being trained to obtain the soft class label estimates. It is demonstrated that substantial gains in classification performance may be achieved from the use of the approach when the labelled data do not represent adequately the entire class distributions. The experimental investigations performed have shown that the approach proposed may be successfully used to train neural networks for learning different classification problems.