A novel information geometric approach to variable selection in MLP networks