Preprocessing of Training Set for Back Propagation Algorithm : Histogram Equalization

This paper considered a preprocessing of input data to improve the performance of standardized BP learning algorithm. In general, even a simple uniform transformation into a proper range improves the efficiency of the BP algorithm. However, uniform transformation becomes inefficient if the input distribution is very skewed. To correct this problem, we propose a modified histogram equalization technique that converts the skewed and irregular distribution to a distribution favorable to the BP algorithm. Our simulation with the S&L failure classification data indicates that the proposed preprocessing leads to a better performance of the BP algorithm than a simple transformation or the conventional histogram equalization