A new pruning algorithm for neural network dimension analysis

The choice of network dimension is a fundamental issue in neural network applications. An optimal neural network topology not only reduces the computational complexity, but also improves its generalization capacity. In this research, a new pruning algorithm based on cross validation and sensitivity analysis is developed and compared with three existing pruning algorithms on various pattern classification problems. Computer simulation results show the network size can be significantly reduced using this new algorithm while the neural network still maintains satisfactory generalization accuracy.