基于敏感性的带有结构调整的Madaline学习算法

How to design a proper architecture for solving a given problem is an important issue in neural network research. The existing training algorithms usually focus on improving training accuracy by only adjusting a neural network’s weights, and few of them adaptively adjust the network’s architecture. However, the architecture is indeed very critical for training neural networks to have high performance and needs to be coped with in the training process. In this paper, we present a new training algorithm of Madalines, which takes architecture adjusting into consideration. The algorithm can train Madalines with smaller architectures and higher generalization performance. Experimental results have demonstrated the effectiveness of the algorithm.