One method is to drop variables with small coefficients in the top n principal components. This is feature selection, by your definition.