Full text
| |
| Author(s): |
Luís Augusto Martins Pereira
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2018-07-27 |
| Examining board members: |
Ricardo da Silva Torres;
Hélio Pedrini;
Alexandre Luís Magalhães Levada;
Aparecido Nilceu Marana;
Marco Antonio Garcia de Carvalho;
Roberto de Alencar Lotufo
|
| Advisor: | Ricardo da Silva Torres |
| Abstract | |
Domain shift is a phenomenon observed when two related domains ¿ a source (training set) and a target domain (test set) ¿ have a mismatch between their marginal probability distributions. This phenomenon is prevalent in machine learning, particularly in real-world applications. In computer vision, for instance, domain shift occurs essentially because visual data are often captured by different devices and under varied imaging conditions, such as scene, pose, and illumination. Under domain shift, however, conventional classifiers often fail to achieve desirable performances at test time. This is because they assume a stationary environment, that is, source and target domain are supposedly drawn from the same probability distribution. How to overcome the limitation imposed by this assumption is the question that guides this study. Here, I present algorithmic mechanisms to approach the domain-shift problem, thereby leading to an improvement on classification performances. The algorithms were proposed under the domain adaptation semi-supervised paradigm, in which a fully labeled source data and a partially labeled target data are available to guide the domain adaptation. The domain-shift problem is addressed by these algorithms in two different fashion: i) by learning a new invariant feature representation across domains; and ii) by combining a new feature representation with a based-domain-adaptation model. In this context, I found that semi-supervised subspace feature learning constrained by semantic links across domains ¿ namely inter-domain pairwise constraints ¿ is an effective mechanism to reduce the domain shift. Additionally, I found that combining the learning of invariant features in a Reproducing Kernel Hilbert Space with a max-margin domain-adaptation model yields a semi-supervised method that reduces the domain shift more than these solutions separately (AU) | |
| FAPESP's process: | 15/09169-3 - Domain adaptation with minimal supervision in multimedia problems |
| Grantee: | Luis Augusto Martins Pereira |
| Support Opportunities: | Scholarships in Brazil - Doctorate |