NEURAL NETWORK TRAINING BY MEANS OF SPECIALIZED LOSS FUNCTIONS Russian patent published in 2019 - IPC G06N3/08 

FIELD: education.

SUBSTANCE: group of inventions relates to the training of a neural network by using specialized loss functions. Method comprises: obtaining, by a computer system, a set of training data containing a plurality of images, where each image from a set of training data is associated with a class identifier from a set of classes; computing, using a neural network, a plurality of feature vectors, wherein each feature vector from the plurality of feature vectors represents an image from a set of training data in the image feature space; calculating, for a set of training data, a loss function value representative of a plurality of probabilities, where each probability from a plurality of probabilities represents a hypothesis associating an image from the set of training data with a class associated with said image in accordance with a set of training data, wherein the loss function further displays a plurality of distances, where each distance from the plurality of distances is calculated in the image feature space between the feature vector representing the image from the set of training data, and the center of the class associated with the image in accordance with the set of training data; and tuning the neural network parameter based on the loss function value.

EFFECT: high quality and efficiency of optical character recognition.

20 cl, 7 dwg