A Structural Optimization Algorithm for Complex-Valued Neural Networks
Zhongying Dong, He Huang · 2019
A suitable network structure can not only save the computing resources but also enhance the generalization ability of the constructed neural network. A pruning algorithm with group lasso regularization is proposed in this paper, where both the superfluous hidden and input neurons can be efficiently removed. Based on this algorithm, appropriate complex-valued neural networks (CVNNs) are designed for practical applications. It is noted that the numerical ill-posed problem would be caused by the nondifferentiability of the group lasso regularization at the origin. To overcome it, a smooth function is introduced to approximate the regularization term. Then, CVNNs with good generalization capability and reasonable network structure are obtained. Experimental results on some benchmark classification problems are provided to show the performance of the proposed pruning algorithm.