Abstract:

To reduce the dimension of a neural network and improve the generalization capability of the extreme learning machine (ELM) network, Akaike information criterion (AIC) was implemented to choose a suitable number of hidden units, and the modified Gram-Schmidt (MGS) method was also implemented to automatically adjust the network parameters. In comparison with the conventional ELM learning method on several commonly used regressor benchmark problems, the improved ELM algorithm could achieve  a  compact network with much faster training speed and satisfactory accuracy.

Key words: extreme learning machine, Akaike information criterion, modified Gram-Schmidt algorithm, feedforward neural network