论文信息 - Effect of initial values in simple perception

Effect of initial values in simple perception

The initial value of the network weight vector can contain a large amount of information. One of the possibilities to use this information is to utilise a weighted combination of initial and final values of the weights. An alternative is the early stopping. To utilise an information contained in the initial weight vector W/spl circ//sub 0/ one needs to know the accuracy of determination of W/spl circ//sub 0/ and W/spl circ//sub 1/, a final weight. The analytical expression of an optimal weighting factor greatly depends on the parameter estimation schema as well as on the learning-set size. Utilisation of the additional validation-set is a universal method.

S. Raudys | S.-I. Amari

[1] Thierry Denoeux,et al. Initializing back propagation networks with prototypes , 1993, Neural Networks.

[2] Sarunas Raudys,et al. On Dimensionality, Sample Size, and Classification Error of Nonparametric Linear Classification Algorithms , 1997, IEEE Trans. Pattern Anal. Mach. Intell..

[3] Sarunas Raudys,et al. Evolution and generalization of a single neurone: : II. Complexity of statistical classifiers and sample size considerations , 1998, Neural Networks.

[4] Shun-ichi Amari,et al. A Theory of Adaptive Pattern Classifiers , 1967, IEEE Trans. Electron. Comput..

[5] Shun-ichi Amari,et al. Four Types of Learning Curves , 1992, Neural Computation.

[6] M. Skurikhina,et al. The role of the number of training samples on weight initialisation of artificial neural net classifier , 1992, [Proceedings] 1992 RNNS/IEEE Symposium on Neuroinformatics and Neurocomputers.