Title :
On the speed of training networks with correlated features
Author :
Bakker, Robert R N ; Kraaijveld, Martin A. ; Duin, Robert P W ; Schmidt, Wouter F.
Author_Institution :
Dept. of Appl. Phys., Delft Univ. of Technol., Netherlands
Abstract :
The learning speed of the adaptive linear combiner is determined by the condition number of the input correlation matrix of the training data. With known properties of such correlation matrices, it is shown that increasing the dimensionality of the feature space of an adaptive linear combiner will never increase its learning speed. In fact, the learning speed will at best remain equal, but will deteriorate in most cases
Keywords :
learning (artificial intelligence); neural nets; adaptive linear combiner; condition number; correlated features; dimensionality; feature space; input correlation matrix; learning speed; training networks; Adaptive signal processing; Automatic logic units; Eigenvalues and eigenfunctions; Learning systems; Least squares approximation; Pattern recognition; Physics; Time factors; Training data; Vectors;
Conference_Titel :
Neural Networks, 1993., IEEE International Conference on
Conference_Location :
San Francisco, CA
Print_ISBN :
0-7803-0999-5
DOI :
10.1109/ICNN.1993.298680
