Collaborative image classification for expansible datasets

• Ano de defesa: 2024
• Autor(a) principal: Padilha, Bruno
• 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: Big data; Causal inference; Composição de invariantes; Conjuntos de dados expansíveis; Deep learning; Expansive datasets; Generalização; Generalization; Inferência causal; Invariant composition
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-25082025-112227/

During the last decade, we have been experiencing a striking evolution of Deep Learning models along with our capacity of generating and storing data. Nowadays, it is common to find application domains generating data streams, subjecting models trained on static datasets to knowledge obsolescence. Several authors are proposing learning methods based on invariant discovery inspired on causal inference. In this work, we assume the following hypothesis: the generalization of Deep Learning models for image classification is correlated to its ability to precisely capture Class Invariants in expansible datasets. In order to validate this hypothesis, we propose a 5-Step learning strategy to approximate global invariants by composition of local invariants learned from subdomain segmented data of expansible datasets. Our experiments conducted with raw video data from USP-EMS, for the classes cyclist, motorcyclist and pedestrian, showed how the domain segmentation and local invariants composition to approximate global invariants through knowledge sharing is indeed a more suitable approach to overcome generalization challenges in expansible datasets. Additionally, we are developing our experimental software framework to manage subdomain segmentation, datasets generation and computational resources to expedite parallel training in distributed environments. Next steps include scalability experiments for thousands of cameras with increased number of classes, and learning of object interactions with temporal dimension to support suspect behavior analysis with expansible datasets in multi-camera environments.