We propose a new criterion for identifying the so-called "neglected" classes during the training of DNNs, i.e. the classes which stop to optimize early in the training procedure. Moreover, based on ...