A New Multilayer Perceptron Pruning Algorithm for Classification and Regression Applications

Optimizing the structure of neural networks remains a hard task. If too small, the architecture does not allow for proper learning from the data, whereas if the structure is too large, learning leads to the well-known overfitting problem. This paper considers this issue, and proposes a new pruning approach to determine the optimal structure. Our algorithm is based on variance sensitivity analysis, and prunes the different types of unit (hidden neurons, inputs, and weights) sequentially. The stop criterion is based on a performance evaluation of the network results from both the learning and validation datasets. Four variants of this algorithm are proposed. These variants use two different estimators of the variance. They are tested and compared with four classical algorithms on three classification and three regression problems. The results show that the proposed algorithms outperform the classical approaches in terms of both computational time and accuracy.