On-Target AdaptationDownload PDF

Published: 28 Jan 2022, Last Modified: 13 Feb 2023ICLR 2022 SubmittedReaders: Everyone
Keywords: domain adaptation, source-free adaptation, unsupervised domain adaptation
Abstract: Domain adaptation seeks to mitigate the shift between training on the source data and testing on the target data. Most adaptation methods rely on the source data by joint optimization over source and target. Source-free methods replace the source data with source parameters by fine-tuning the model on target. Either way, the majority of the parameter updates for the model representation and the classifier are derived from the source, and not the target. However, target accuracy is the goal, and so we argue for optimizing as much as possible on target. We show significant improvement by on-target adaptation, which learns the representation purely on target data, with only source predictions for supervision (without source data or parameter fine-tuning). In the long-tailed classification setting, we demonstrate on-target class distribution learning, which learns the (im)balance of classes on target data. On-target adaptation achieves state-of-the-art accuracy and computational efficiency on VisDA-C and ImageNet-Sketch. Learning more on target can deliver better models for target.
One-sentence Summary: Because target accuracy is the goal, we argue for optimizing as much as possible on the target data, and decouple the adaptation of source predictions from the source representation to do so.
Supplementary Material: zip
10 Replies

Loading