Features transfer learning between domains for image and video recognition tasks

• Ano de defesa: 2020
• Autor(a) principal: Santos, Fernando Pereira dos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Alinhamento de variedades; Aprendizado profundo; Cross-domain; Cruzamento de domínios; Deep learning; Features transfer learning; Generalization measures; Manifold alignment; Medidas de generalização; Transferência de aprendizado de características
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-19032020-083537/

Feature transfer learning aims to reuse knowledge previously acquired in some source dataset to apply it in another target data and/or task. A requirement for the transfer of knowledge is the quality of feature spaces obtained, in which deep learning methods are widely applied since those provide discriminative and general descriptors. In this context, the main questions include: what to transfer align the data distribution from source and target, and adjusting the parameters to increase the models generalization capability; how to transfer investigating methods that work on the features spaces or also on the learned models; and when to transfer studying which datasets are mode adequate for transferring, considering discrepancies between source and target data, such as they different acquisition settings, clutter and illumination variation, among others. This thesis advocates that the focus should be in transferring feature spaces, learned by convolutional neural networks, in particular investigating the descriptive potential of inner and initial layers of such deep convolutional networks, and the approximation of feature spaces before aligning the data distribution in order to allow for better solutions, as well as the use of both labeled and unlabeled for feature learning. Besides the transfer learning methods, such as fine-tuning and manifold alignment, with use of classical evaluation metrics for recognition performance, a generalization metric between domains is also proposed to evaluate transfer learning. This thesis contributes with: an analysis of multiple descriptors contained in supervised deep networks; a new architecture with a loss function for semi-supervised deep networks (Weighted Label Loss), in which all available data, labeled or unlabeled, are incorporated to provide learning; and a new generalization metric (Cross-domain Feature Space Generalization Measure) that can be applied to any model and evaluation system