Joint Weakly Parameter-Shared and Higher Order Statistical Criteria for Domain Adaptation
| Autor: | Guangrong Bian, Pan Wang, Ming Zhao |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
Artificial neural network
Computer science business.industry Feature extraction Pattern recognition 02 engineering and technology 010501 environmental sciences 01 natural sciences Data modeling 0202 electrical engineering electronic engineering information engineering Task analysis Probability distribution 020201 artificial intelligence & image processing Central moment Artificial intelligence Transfer of learning business Classifier (UML) 0105 earth and related environmental sciences |
| Zdroj: | 2017 International Conference on Industrial Informatics - Computing Technology, Intelligent Technology, Industrial Information Integration (ICIICII). |
| DOI: | 10.1109/iciicii.2017.41 |
| Popis: | Domain adaptation is a particular case of transfer learning that applies labeled data in one or more related source domains, to learn a classifier for unseen or unlabeled data in a target domain. Recent studies reveal that a deep neural network can learn transferable features which generalize well to novel tasks with domain adaptation. However, most existing methods use the strongly parameter-shared architecture with the same weights for both source and target data, and transferring criteria commonly uses maximum mean discrepancy (MMD). In order to get more flexible and robust features, a new architecture joint weakly parameter-shared structure and higher order statistical criteria as transferring criteria is introduced. This architecture is a two-stream architecture, one of which operates in the source domain and the other in the target domain. Different from the strongly parameter-shared structure, the weakly parameter-shared structure allow the use of separate layers for different domain. In addition, central moment discrepancy (CMD), as a higher order statistical criteria, is used in the approach to measure domain feature discrepancy more accurate. The approach can extract more appropriate domain-invariant features and domain-specific features, making our model more robust and accurate. Experimental results on the digital datasets will show its competitive performance as compared with the other state-of-the-art methods. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|