Analisis Pengawalan Pemberat Rangkaian Neural Perambatan Balik untuk Pengecaman Aksara Jawi

Abstract

One of the factors that influences the recognition ability of a neural network is the initial values given to the weight vector during the training phase. The network may be trapped into a local minima if the initial weights are not chosen carefully. This paper presents an analysis of the ability of the network to recognise Jawi characters after it was trained using different methods of weight initialization. Three most common methods are zero, random and Nguyen-Widrow random. This paper presents the effect of these three methods on the ability of the network's recognition.