Recent Research Trends in Unsupervised Adaptation Techniques for Data Changes

Abstract

Changes in data distribution over time are one of the main causes to degrade performance of AI systems using machine learning models trained beforehand. We have introduced general and recently published methods for detecting and adapting such changes in real time without labeling data by hand at JSAI2020 and JSAI2021, respectively. In future operations in AI systems, there are growing needs for new adaptation techniques that do not depend on source data used to train machine learning models before adaptation from the viewpoint of data privacy and its portability, and that can respond to various kinds of changes occurred in input data. In this paper, we introduce the latest research trends in unsupervised adaptation techniques to data changes, focusing on unsupervised concept drift adaptation and unsupervised domain adaptation methods presented at major international conferences in the field of machine learning held in recent years, especially after 2019. In addition, we will discuss the effectiveness of these methods in solving new challenges in AI operations, such as preserving data privacy.