Variability in generalisation curves and the effects of linear scaling thereon
Alan M. Parkinson · 2003
It is demonstrated that different linear scalings of input data can have significant effects on stability of the learning trajectory. Using a feed forward network with sigmoid output function, two different financial data sets were trained under varying conditions. It was found that a range of 0.3-0.7 gave much more consistent results than the commonly employed 0.1-0.9. The variability was shown to have two causes, one of which was an artefact of presentation sequence.