正負の事例の提示頻度の逆伝搬学習に与える影響

概要

論文の詳細を見る
For the error back propagation learning through single-output three layer networks, this paper points out that, although every example updates the network parameters to surely decrease its own error, some examples other than that update them favorably to decrease it but some unfavorably to increase it. Whether it becomes decreased or increased is uniquely determined in terms of the example led to the parameter change and the interested example. These increments and decrements due to the other examples can be shown to be relaxed on an average by presenting and learning both positive and negative examples in the same frequency. Based on this observation, such a method is considered that, when the positive and negative examples are different in number, both examples are presented and learned in the same frequency by overlapping the fewer ones. By simulations, this method is shown to have a possibility to improve the learning convergency.
近畿大学の論文
1994-12-01