Domain generalization and domain adaptive learning method based on data expansion consistency

Abstract

The invention belongs to the field of artificial intelligence machine learning, and discloses a domain generalization and domain self-adaption method based on data expansion consistency, which comprises the following steps: S1, according to task requirements, designing a prediction model p theta (y | x) based on a deep neural network, theta being a model parameter, and model output being conditional probability distribution of marking y under the condition of a given sample x; S2, constructing a data expander according to task characteristics, converting the sample, and keeping the core content of the sample unchanged, so as to keep the real mark of the converted sample unchanged; S3, constructing a multi-task loss function consisting of supervised loss and data expansion consistency lossby utilizing the original training sample and the expanded sample, and training to obtain p theta * (y | x); and S4, applying the trained model p theta * (y | x) to a target field test sample for prediction. The domain generalization and domain adaptive learning method is simple, universal and good in performance, and the technical problem of domain offset can be better solved.

Description

technical field [0001] The invention belongs to the field of machine learning of artificial intelligence, and specifically relates to domain generalization (Domain Generalization) and domain adaptation (Domain Adaptation) deep learning. Background technique [0002] In recent years, deep learning has achieved very successful applications in the fields of image, text and speech, but deep learning still faces the problem of domain shift (domain shift), that is, when the test data (target domain) and training data (source domain) When the distribution of σ is inconsistent, the model testing performance will drop sharply. Domain generalization and domain adaptive learning are machine learning methods that aim to reduce the performance degradation caused by domain shift. Domain adaptive learning learns a model suitable for target domain data through labeled source domain samples and unlabeled (or a small amount of labeled) target domain data. Domain adaptive learning has certai...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to solve the problems of the current domain generalization and domain self-adaptation methods, such as complex models and difficult training, and the need to introduce auxiliary task networks and parameters, etc., and propose a simple, general, and good performance domain generalization and domain self-adaptive learning method , to better solve the technical problems of domain offset

Images