GANDA : generative adversarial network-based domain adaptationGANDA : 생산적 적대 신경망기반 도메인 적응 기법

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 174
  • Download : 0
Being trained with a large number of labeled data, deep neural networks can still fail to generalize to a target domain whose distribution is different from the source data distribution by a domain shift. On the other hand, labeling data to achieve an appropriate accuracy on the target domain is expensive and sometimes impracticable. Unsupervised domain adaptation addresses this problem by learning a model aiming at target dataset using labeled examples from source domain and unlabeled examples from target domain. The two key requirements to achieve successful adaptation are domain-invariant features extraction and discriminative features extraction. Domain alignment method extracts the domain-invariant features by matching the feature distribution of source and target domain. However, the extracted domain-invariant features are not guaranteed to be discriminative. To solve this problem, we propose Generative Adversarial Network-based Domain Adaptation (GANDA) method to extract the discriminative features. This method is orthogonal to the existing domain adaptation works and can be added as an additional component to further improve the performance of the models. Being evaluated with comprehensive experiments, our method outperforms or approaches the state-of-the-art methods on various standard domain adaptation tasks.
Advisors
Yi, Yungresearcher이융researcher
Description
한국과학기술원 :전기및전자공학부,
Publisher
한국과학기술원
Issue Date
2019
Identifier
325007
Language
eng
Description

학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2019.8,[iii, 27 p. :]

Keywords

Unsupervised domain adaptation▼adeep learning▼agenerative adversarial network; 관리되지 않는 도메인 적응▼a깊은 학습▼a생성 적 적군 네트워크▼a관리되지 않는 도메인 적응▼a관리되지 않는 도메인 적응

URI
http://hdl.handle.net/10203/283066
Link
http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=875360&flag=dissertation
Appears in Collection
EE-Theses_Master(석사논문)
Files in This Item
There are no files associated with this item.

qr_code

  • mendeley

    citeulike


rss_1.0 rss_2.0 atom_1.0