Evaluating the Informativity of a Training Sample for Image Classification by Deep Learning Methods