File(s) under permanent embargo
Source-Free Implicit Semantic Augmentation for Domain Adaptation
Unsupervised Domain Adaptation (UDA) challenges the problem of alleviating the effect of domain shift. Common UDA methods all require labelled source samples. However, in some real application scenarios, such as Federated Learning, the source data is inaccessible due to data privacy or intellectual property, and only a pre-trained source model and target data without labels are accessible. This challenging problem is called Source-Free Domain Adaptation (SFDA). To address this, we introduce a generation encoder to generate source prototypes depending on the hidden knowledge from the pre-trained source classifier. The generated source prototypes can describe the distribution of source samples in the feature space to a certain extent to solve the Source-Free problem. We also propose Source-Free Implicit Semantic Augmentation (SFISA) for adaptation. SFISA consists of two main stages: source and target class prototypes generation and Source-Free semantic augmentation adaptation based on generated class prototypes. Extensive experiments on the UDA benchmarks demonstrate the efficacy of our generation encoder and augmentation method SFISA.
History
Volume
13630 LNCSPagination
17-31Publisher DOI
ISSN
0302-9743eISSN
1611-3349ISBN-13
9783031208645Publication classification
E1.1 Full written paper - refereedTitle of proceedings
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)Publisher
Springer Nature SwitzerlandSeries
Lecture Notes in Computer ScienceUsage metrics
Categories
No categories selectedKeywords
Licence
Exports
RefWorks
BibTeX
Ref. manager
Endnote
DataCite
NLM
DC