Metric transfer learning via geometric knowledge embedding
- PDF / 1,545,417 Bytes
- 14 Pages / 595.224 x 790.955 pts Page_size
- 21 Downloads / 241 Views
Metric transfer learning via geometric knowledge embedding Mahya Ahmadvand1 · Jafar Tahmoresnezhad1
© Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract The usefulness of metric learning in image classification has been proven and has attracted increasing attention in recent research. In conventional metric learning, it is assumed that the source and target instances are distributed identically, however, real-world problems may not have such an assumption. Therefore, for better classifying, we need abundant labeled images, which are inaccessible due to the high cost of labeling. In this way, the knowledge transfer could be utilized. In this paper, we present a metric transfer learning approach entitled as “Metric Transfer Learning via Geometric Knowledge Embedding (MTL-GKE)” to actuate metric learning in transfer learning. Specifically, we learn two projection matrices for each domain to project the source and target domains to a new feature space. In the new shared sub-space, Mahalanobis distance metric is learned to maximize inter-class and minimize intra-class distances in target domain, while a novel instance reweighting scheme based on the graph optimization is applied, simultaneously, to employ the weights of source samples for distribution matching. The results of different experiments on several datasets on object and handwriting recognition tasks indicate the effectiveness of the proposed MTL-GKE compared to other state-of-the-arts methods. Keywords Metric learning · Transfer learning · Geometric knowledge embedding · Mahalanobis distance metric
1 Introduction Today, variety of web technologies, social media and digital devices continuously generate enormous amount of visual data such as images and videos in increasing manner [1].This case confront us with one of the challenging subjects of big data problem such as data management, in the rising stream of novel applications and corresponding data generation. A prerequisite for big data management is labeling and classification of existing data. However, the researchers confront with an entirely sparse labeled data, which is not enough for training an accurate classifier. On the other hand, labeling this enormous amount of data may require an expert to use an expensive way. In such circumstances, transfer learning and domain adaptation methods [2, 3] can be used to utilize previous source labeled data to create a classifier and apply it on new task in target data. Conventional machine learning
Jafar Tahmoresnezhad
[email protected] 1
Faculty of IT, Computer Engineering, Urmia University of Technology, Urmia, Iran
methods have the assumption of same distribution of source and target samples, while this assumption in real world application is not considered anymore and the cross-domain problem arises. Hereupon, domain adaptation with the aim of reducing the destructive impact of cross domain problem on the classification accuracy and learning a domain invariant model from training data, is introduced. However, state-of-
Data Loading...
