Method and device for unsupervised domain adaptation

Abstract

The invention discloses an unsupervised domain adaptation method, device and electronic equipment; the method includes: using a feature extractor to obtain the first eigenvalue of the source domain and the first eigenvalue of the target domain, and using a preset mean value algorithm to obtain the target Domain pseudo-label; use the output of the feature extractor and classifier to the source domain to establish a source domain classification loss function; use the first eigenvalue of the target domain and the centroid of the target domain, and the second eigenvalue of the target domain to establish an overall Clustering loss function; use the output of the feature extractor to the source domain and the target domain, and the output of the classifier to the source domain and the target domain to establish a conditional distribution overall loss function; use the above loss function to establish an overall loss function, and The fit of the supervised domain adaptation network is evaluated to update the parameters of the unsupervised domain adaptation network.

Description

technical field [0001] Embodiments of the present disclosure relate to the technical field of deep learning, and in particular to a method and device for unsupervised domain adaptation. Background technique [0002] Unsupervised domain adaptation aims to find an efficient model for an unlabeled target domain by exploiting the knowledge of the labeled source domain, where the data of the source and target domains are different but related. The existing unsupervised domain adaptation methods often ignore the potential discriminative features and conditional distribution differences of the data in the target domain, that is to say, the existing unsupervised domain adaptation techniques cannot achieve domain invariance at the same time. Mining of discriminative features, and alignment of class features. [0003] Based on this, there is a need for a scheme that can naturally unify the mining of domain-invariant discriminative features and the alignment of class features into a s...

AI Technical Summary

Problems solved by technology

The existing unsupervised domain adaptation methods often ignore the potential discriminative features and conditional distribution differences of the data in the target domain, that is to say, the existing unsupervised domain adaptation techniques cannot achieve domain invariance at the same time. Mining of discriminative features, and alignment of class features

Images