Analysis on the Convergence of Dyadic Wavelet Based Neural Network with Varying Learning Rate and Resolution for Function Learning

This paper presents the analysis of results on the generalisation of dyadic based wavelet neural network which are trained with uniform distribution from input space. The focus is to mainly quantify the significance of learning rate and the resolution so as to ensure an acceptable generalization accuracy for function learning simulations. The proposed network is based on orthonormal basis functions and trained with stochastic gradient algrothim. The simulations of developed dyadic wavelet based architecture and its learning algorithm justifies the effectiveness of the scaling function characteristics. Experimental results reveal that training and tuning the various simulation parameters of the network and its properties has greater influence on the generalization and convergence ability of the Dyadic Wavelet Neural Network (DWNN).