Combining multi-layer perceptrons in classification problems

摘要

The impredictability of the performance after the training often limits the use of the Multi Layer Perceptron(MLP). The tuning of the gain parameter is committed to heuristic techniques and therefore these networks often tend to local minima. The choice of the training set can strongly influence the generalization ability producting overfitting and underfitting phoenomena. This paper presents a possible soltion to these problems: the use of more networks to overcome the inefficiency of the single MLP. We experimented some techniques to merge the outputs of a committee of MLPs and we obtained a global performance free from the above problems. These thchniques were used to classify handwritten digits from the NIST database, obtaining the best score among systems based only on the official database.