METHOD OF TRAINING DEEP NEURAL NETWORKS BASED ON DISTRIBUTIONS OF PAIRWISE SIMILARITY MEASURES

摘要

FIELD: physics.;SUBSTANCE: marked training sample is obtained, where each element of the training sample has a label of the class, to which it belongs; a set of disjoint random subsets of the training sample of input data is formed for a deep neural network in such a way that when combined they represent a training sample; each formed subset of the training sample is transmitted to the input of the deep neural network, obtaining a deep representation of the given subset of the training sample at the output; all pairwise similarity measures between the deep representations of the elements of each subset obtained at the previous step are defined; the similarity measures between the elements that have the same class labels, defined in the previous step, are referred to the similarity measures for positive pairs, and the similarity measures between the elements that have different class labels are referred to the similarity measures for negative pairs; the probability distribution of the values of similarity measures for positive pairs and the probability distribution of the values of similarity measures for negative pairs are determined by using a histogram; a loss function is formed based on the probability distributions of similarity measures defined in the previous step for positive pairs and negative pairs; the generated function in the previous loss step is minimized using the error back propagation method.;EFFECT: increasing the accuracy of training and reducing the time required to adjust the learning parameters for deep views of input data.;10 cl, 7 dwg