A transductive transfer learning approach for image classification
- PDF / 2,187,873 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 105 Downloads / 223 Views
ORIGINAL ARTICLE
A transductive transfer learning approach for image classification Samaneh Rezaei1 · Jafar Tahmoresnezhad1 · Vahid Solouk1 Received: 25 May 2019 / Accepted: 8 September 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Among machine learning paradigms, unsupervised transductive transfer learning is useful when no labeled data from the target domain are available at training time, but there is accessible unlabeled target data during training phase instead. The current paper proposes a novel unsupervised transductive transfer learning method to find the specific and shared features across the source and the target domains. The proposed learning method then maps both domains into the respective subspaces with minimum marginal and conditional distribution divergences. It is shown that the discriminative learning across domains leads to boost the model performance. Hence, the proposed method discriminates the classes of both domains via maximizing the distance between each sample-pairs with different labels and via minimizing the distance between each instance-pairs of the same classes. We verified our approach using standard visual benchmarks, with the average accuracy of 46 experiments as 76.5%, which rates rather high in comparison with other state-of-the-art transfer learning methods through various cross-domain tasks. Keywords Machine learning · Unsupervised transfer learning · Cross-domain problems · Discriminative learning · Respective subspaces
1 Introduction Transfer learning has been the interest of many researches for the incurring performance boost of learning in target domain, which is originated from inheriting well-learned knowledge of source domain. The transductive transfer learning exploits the labeled training set and unlabeled test set for training the model to infer the labels of unlabeled test set [1]. For a new sample, the transductive transfer algorithm trains the model on entire data including even the new sample. For an example in biological sequence classification, the forthcoming unlabeled samples with different feature distribution needs to be labeled according to previous experiments [2].
* Jafar Tahmoresnezhad [email protected] Samaneh Rezaei [email protected] Vahid Solouk [email protected] 1
Faculty of Information Technology and Computer Engineering, Urmia University of Technology, Urmia, Iran
In order to reduce the distribution difference across domains, transfer learning uses the following three lines of strategies. Model-based methods train a model with source domain and adapt the parameters of model for target domain [3, 4]. Instance-based methods re-weight the source samples and train a model on source data to adapt with target domain [5]. Feature-based methods aim to find feature sub-spaces where the distribution divergence across domains is minimized [6–12]. Feature-based domain adaptation methods, based on the type of features in latent space are categorized into two strategies including data-alignment and su
Data Loading...
