Deakin University
Browse

File(s) under permanent embargo

Source-Free Implicit Semantic Augmentation for Domain Adaptation

conference contribution
posted on 2023-02-23, 02:41 authored by Z Zhang, Zili ZhangZili Zhang
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 LNCS

Pagination

17-31

ISSN

0302-9743

eISSN

1611-3349

ISBN-13

9783031208645

Publication classification

E1.1 Full written paper - refereed

Title of proceedings

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Publisher

Springer Nature Switzerland

Series

Lecture Notes in Computer Science

Usage metrics

    Research Publications

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC