Rethinking Effectiveness of Unsupervised Domain Adaptation Methods
Keywords: domain adaptation, transfer learning, unsupervised learning
TL;DR: An empirical analysis of efficiency of unsupervised domain adaptation methods through the lens of various modeling choices.
Abstract: Recently, significant progress has been made in unsupervised domain adaptation (UDA) through techniques that enable reduction of the domain gap between labeled source domain data and unlabeled target domain data. In this work, we examine the diverse factors that may influence the effectiveness of UDA methods, and devise a comprehensive empirical study through the lens of backbone architectures, quantity of data and pre-training datasets to gain insights into the effectiveness of modern adaptation approaches on standard UDA benchmarks. Our findings reveal several non-trivial, yet valuable observations: (i) the benefits of adaptation methods decrease with advanced backbones, (ii) current methods under-utilize unlabeled data, and (iii) pre-training data matters for downstream adaptation in both supervised and self-supervised settings. To standardize evaluation across various UDA methods, we develop a novel PyTorch framework for domain adaptation and will release the framework, along with the trained models, publicly.
Supplementary Material: zip
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1000
Loading