TRAINING OF DNN-STUDENT BY MEANS OF OUTPUT DISTRIBUTION

摘要

FIELD: computer equipment.;SUBSTANCE: invention relates to the field of machine learning algorithms. To generate a DNN classifier by "learning" the DNN-student model based on a larger, more accurate DNN teacher model, a DNN student can be trained on the basis of unmarked training data by passing unmarked training data through a DNN teacher, which can be trained based on tagged data. Iterative process is used to train a DNN student by minimizing the divergence of the pin assignments based on the DNN teacher and student models. For each iteration before convergence, the difference in the outputs of these two DNNs is used to update the DNN student model, and the findings are determined again, using unmarked training data.;EFFECT: technical result is an increase in the accuracy of the DNN (Deep Neural Network) model with a reduced size.;10 cl, 7 dwg