Proposta de Tese de Matheus Vinícius Todescato

Proposta de Tese de Doutorado

Aluno(a): Matheus Vinícius Todescato
Orientador(a): Joel Luis Carbonera

Título: Learning Under Mixed Distributions in Label-Scarce Scenarios: OOD-Aware Semi-Supervised Image Classification
Linha de Pesquisa: Aprendizado de Máquina, Representação de Conhecimento e Raciocínio

Data: 11/12/2025
Hora: 10:30
Local: Esta banca ocorrerá de forma remota. Acesso público disponibilizado pelo link https://mconf.ufrgs.br/webconf/00179534.

Banca Examinadora:
-Thiago Lopes Trugillo da Silveira (UFSM)
-Cláudio Rosito Jung (UFRGS)
-Adriano Quilião de Oliveira (UFSM)

Presidente da Banca: Joel Luis Carbonera

Resumo: In recent years, deep learning models have achieved remarkable success in several computer vision tasks, but their effectiveness in domain-specific scenarios remains constrained by the scarcity of labeled data in various domains. Since manual labeling is costly and time-consuming, large amounts of unlabeled data remain underutilized, reinforcing the need for approaches that can leverage both labeled and unlabeled samples. Semi-supervised (SSL) methods are usually the solution, yet in practice, unlabeled data often contain out-of-distribution (OOD) samples, which can severely degrade SSL performance. This challenge results in difficulties for training vision models in practical environments with low-labeled data and unlabeled data that is a mixture of in-distribution and OOD instances. To address this issue, this work proposes a lightweight and modular pipeline that integrates OOD detection with SSL to enable effective training of ima ge classification models. The combination is designed to mitigate error propagation, maximize the utility of limited labeled examples, and improve robustness in domain-specific contexts. In this proposal, we present and discuss the research problem, outline the central hypothesis, and introduce a preliminary OOD detection approach capable of operating in zero-shot and few-shot settings. Finally, we discuss the expected impact of the method, highlight the remaining challenges for full pipeline integration, and define the next steps required to advance toward enabling reliable training of vision models in scenarios where annotation costs and distribution mismatch remain key obstacles to the adoption of AI.

Palavras-Chave: Image Classification,OOD detection,Semi-Supervised Learning,Label-Scarce,Transfer Learning.