透過目標整合的表徵學習來根除通用域適應中的隱性不匹配
Other Title
Uprooting Implicit Misalignment in Universal Domain Adaptation by Target-Integrated Representation Learning
Journal
2024臺大學士論文獎
Date Issued
2024
Author(s)
Hung-Chieh Fang
Advisor
Abstract
Universal Domain Adaptation (UniDA) seeks to address the challenges posed by distribution shifts between source and target domains without assuming any relationships between their label sets. In this paper, we demonstrate that misalignment in Domain Adaptation (DA) can occur in two scenarios. The first scenario, which we term as implicit misalignment, arises from a source-biased feature space. The second scenario, termed explicit misalignment, occurs when feature alignment methods mistakenly align incorrect features, further exacerbating the misalignment issue. We further demonstrate that UniDA can intensify these two types of misalignment. Previous works have primarily addressed the explicit misalignment issue by designing improved methods for uncertainty measurement to down-weight the influence of private samples in feature alignment. However, these methods often fall short when implicit misalignment is substantial. To address this, we propose a source-label-free target incorporation module that leverages self-supervised learning to mitigate implicit misalignment. Unlike traditional methods, self-supervised learning promotes a broader diversity of guidance, reducing the likelihood of incorrect feature alignments. Our experimental results on the Office-Home and Office31 dataset with different settings indicate that our approach is more effective at handling problems with large implicit misalignment.
Subjects
Domain Adaptation
Universal Domain Adaptation
Machine Learning
Self-supervised Learning
Publisher
國立臺灣大學資訊工程學系
Description
指導教授:林軒田
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
資工系_方泓傑(已保全).pdf
Size
1.28 MB
Format
Adobe PDF
Checksum
(MD5):d02b74fa07bb06551fa8191dc0ef1429