Model training method, data processing method and equipment

Abstract

The embodiment of the invention discloses a model training method, a data processing method and equipment, which can be applied to the field of artificial intelligence, in particular to the field of computer vision. The method comprises the following steps: firstly, establishing a prototype (namely a category center) for each category of each domain (including a target domain and a plurality of source domains), establishing a relation matrix (namely a similarity matrix) based on the calculated prototype in each domain, constructing a target tensor according to the similarity matrix corresponding to each domain, constructing high-order correlation between the domains through the similarity matrix, and fusing the high-order correlation into a target loss function to obtain a target loss function; and enabling the model learning to extract domain-independent features. According to the method, high-order correlation between a plurality of source domains and target domains is mined by means of low-rank constraints of tensors, and the consistency of features extracted from different domains is enhanced, so that the performance of the model is well improved even if the model is in the target domain which is not seen, and the running time and the calculation overhead of a reasoning stage are not additionally increased.

Description

technical field [0001] The present application relates to the field of artificial intelligence, in particular to a model training method, data processing method and equipment. Background technique [0002] With a large amount of labeled data as the "fuel" of network training, deep learning algorithms have shown excellent performance in various computer vision tasks (such as image classification, target detection, video analysis, etc.), machine translation, speech recognition and other fields. . However, there are large inter-domain differences between data from different source domains (i.e., labeled data), and these differences make it straightforward to transfer knowledge learned on a certain source domain to other target domains with different data distributions. The time performance is greatly reduced, which greatly increases the difficulty of deep learning algorithm deployment. [0003] To solve this problem, transfer learning methods, unsupervised domain adaptation (...

AI Technical Summary

Problems solved by technology

[0004] At present, most transfer learning methods mainly reduce the difference between the features extracted from the source domain and the target domain through adversarial training or loss based on adversarial generation. This method learns to extract domain invariant features through adversarial training. The method makes the model training process unstable and increases the difficulty of training; most unsupervised domain adaptation methods are aimed at a single source domain to a single target domain, but in practical applications there are many related source domains that can be used to improve The performance of the model in the target domain; the multi-source domain adaptation algorithm aims to use relevant source domain data to learn domain-independent features common to different source domains related to the task, thereby improving the generalization ability and discrimination of the model in the target domain At present, the multi-source domain adaptation algorithm only considers the pairwise association between the source domain and the target domain, but ignores the association between all domains

Images