Anand R, Mehrotra K G, Mohan C K, Ranka S
Sch. of Comput. and Inf. Sci., Syracuse Univ., NY.
IEEE Trans Neural Netw. 1993;4(6):962-9. doi: 10.1109/72.286891.
The backpropagation algorithm converges very slowly for two-class problems in which most of the exemplars belong to one dominant class. An analysis shows that this occurs because the computed net error gradient vector is dominated by the bigger class so much that the net error for the exemplars in the smaller class increases significantly in the initial iteration. The subsequent rate of convergence of the net error is very low. A modified technique for calculating a direction in weight-space which decreases the error for each class is presented. Using this algorithm, the rate of learning for two-class classification problems is accelerated by an order of magnitude.
对于大多数样本属于一个主导类别的两类问题，反向传播算法收敛非常缓慢。分析表明，出现这种情况的原因是，计算出的网络误差梯度向量在很大程度上由较大的类别主导，以至于在初始迭代中，较小类别中样本的网络误差显著增加。随后网络误差的收敛速度非常低。本文提出了一种在权重空间中计算方向的改进技术，该技术可降低每个类别的误差。使用该算法，两类分类问题的学习速度提高了一个数量级。
