Rethinking Cost-Sensitive Classification in Deep Learning vi