Improving prediction performance of neural networks in pattern classification

Abstract

Neural networks that are especially useful for mapping problems requiring tolerance of some errors and deriving the computing power through their massively parallel-distributed structure have been one of the most efficient approaches for many problems. Moreover, backpropagation is one of the most popular neural networks and is widely applied in various problems. Despite of the many successful applications of backpropagation, it has some drawbacks. One of the serious problems of the backpropagation model is that it is sensitive to the initial value of the weights. The performance in terms of prediction accuracy and computing cost highly depends on the initial weights. Nevertheless, until now, there has been no solution to the significant different performance according to the initial weight configuration. In this article, a prediction rule has been proposed to minimize the effect of initial weights and improve the prediction accuracy on the test data sets. According to the experimental results, a significant improvement of the prediction performance has been achieved by using a proposed prediction rule. The proposed rule could be useful for many other applications of backpropagation to achieve the best performance.

Keywords:

• Neural networks
• Backpropagation
• Initial weights
• Prediction rule