Full text
| |
| Author(s): |
João Gabriel Camacho Presotto
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Rio Claro. 2021-10-22. |
| Institution: | Universidade Estadual Paulista (Unesp). Instituto de Geociências e Ciências Exatas. Rio Claro |
| Defense date: | 2021-08-27 |
| Advisor: | Daniel Carlos Guimarães Pedronette |
| Abstract | |
Despite the impressive recent advances in machine learning techniques, especially in multimedia data understanding, significant challenges remain. One of the main challenges in real-world scenarios is the limited availability of labeled data. In this context, developing methods capable of exploiting the information encoded in the unlabeled data to mitigate the problems associated with insufficient labeled data, and to overcome this issue is something of paramount importance. Weakly supervised learning methods are capable to handle such restrictions by working with estimated or approximate labels as a way to maximize useful training information. In this line of research, we will present two weakly supervised methods that can analyze the relationships between labeled and unlabeled data to expand small labeled training sets. Both use a ranking model and different strategies to examine similarity information encoded in the unlabeled data to identify strong similarity relationships with the labeled data. Such relations will be considered during the training set expansion step. The methods were evaluated in conjunction with different supervised and semi-supervised classifiers, including a recent graph convolutional network. Five different public image datasets were considered with different visual descriptors. Positive accuracy gains were achieved by both methods in the different scenarios when compared to classifiers trained without the aid of our methods and compared to similar expansion techniques, evidencing the strength of both. (AU) | |
| FAPESP's process: | 19/04754-6 - Weakly supervised learning strategies through Rank-based measures |
| Grantee: | João Gabriel Camacho Presotto |
| Support Opportunities: | Scholarships in Brazil - Master |