Training of Deep Neural Networks Based on the Distribution of Pair Similarity Measures

摘要

The present invention relates to training computational systems based on biological models, in particular deep neural networks. The method of training a deep neural network consists of: producing a tagged training sample; Generating a set of non-intersecting any subsets of the training sample of input data; Sending the subset of training samples to an input of a deep neural network in an in-depth representation of the subset of training samples produced at an output; Determining a similarity measure of all pairs between in-depth representations of each subset of elements created in the previous step; Associating the produced deep representation with a similarity measure of positive and negative pairs; Determining their possible distributions using histograms; Generating a loss function based on the possible distribution; Minimizing the loss function with the aid of an error back-propagation method.