Improving neural network learning through output vector scaling
Halis Altun, K.M. Curtis · 2002
We show that through output vector scaling an improvement in the estimation performance of a neural network can be obtained for any training pattern. A smaller estimation error is obtained for the neural network, when it is trained with a sigmoid-like scaled output vector than when it is trained with a linear or nonlinear scaled ones. The most populated region in the output vector domain is determined accordingly through investigation of the nature of the problem to be solved. Through a vector scaling technique in which the nature of the problem is incorporated, one can assign more resolution to a more populated region of the output vector domain. The neural network trained with a set of scaled output vectors results in a more accurate estimation of the output vector, with increased resolution when recall or generalisation is carried out. Results indicate that if the nature of the problem is suitable, the technique can also speed up convergence. The technique is applied to estimate the control parameters of an articulatory speech synthesizer.