Neural network input feature selection using structured l2 − norm penalization

Abstract Artificial neural networks are referred to as universal approximators due their inherent ability reconstruct complex linear and nonlinear output maps conceived input-output relationships from data sets. This can be done by reducing large via regularization in order establish compact models containing fewer parameters aimed at describing vital dependencies In situations where the sets contain non-informative input features, devising a continuous, optimal feature selection technique lead improved prediction or classification. We propose continuous through dimensional reduction mechanism using ‘structured’ l 2 − norm regularization. The implementation is identifying most informative subsets given set an adaptive training mechanism. adaptation involves introducing novel, modified gradient approach during deal with non-differentiability associated of structured penalty. When method applied process sets, results indicate that inputs artificial selected penalization.