Manipulation of Hidden Layers to Improve the Generalization Ability of Neural Networks.

This paper proposed a new algorithm (IM-COH) which was developed based on backpropagation algorithm on neural network (BPNN). This kind of algorithm improved the generalization ability on neural networks to optimize its structure in regularization method. The IM-COH algorithm controls the outputs of nodes in the hidden layers which were manipulated to eliminate the distribution weights problem and enhance the ability to move toward target problems. Furthermore, the IM-COH algorithm maintains the ability to reduce the random noise in the input sample. In this research, the IM-COH algorithm is simulated with two-spiral classification problem and Mackey-Glass time series prediction to compare with Gaussian regularizer (GR) and Laplace regularizer (LR). The results show that the generalization ability of the IM-COH algorithm is better than GR and LR in testing data samples and noise data. In addition, the IM-COH algorithm performs better in solving the two-spiral classification problem comparing to Mackey-Glass time series prediction. [ABSTRACT FROM AUTHOR]