TRAINING OF NEURAL NETWORK MODEL Russian patent published in 2023 - IPC G06N3/08 

FIELD: machine learning.

SUBSTANCE: invention relates to a system, a method, and a data carrier for training of a neural network model. In the method, training data is obtained, which contains data, an annotation for this data, which is determined by a user, and auxiliary data, wherein auxiliary data describes at least one location of interest in data, considered by the user, when determining the annotation for data. The model is trained using training data, including minimization of an auxiliary loss function, which compares at least one location of interest with output data of one or more hidden model layers, and updating of model weights so that to impart increased significance to the specified at least one location of interest in data compared to locations not of interest in data, when annotating data; and minimization of the main loss function, which compares the annotation for data, which is determined by the user, with the annotation created by the model.

EFFECT: increase in a speed and quality of training of a neural network model.

14 cl, 7 dwg