Universidade Federal do Ceará
Mestrado e Doutorado em Ciências da Computação

Título: Joint learning of GP and GPLVM for Semi-Supervised Tasks

Data:22/08/2024
Horário: 16h00
Local:  Sala de Seminários – Bloco 952 e Google Meet (https://meet.google.com/tdu-uemg-kxw)

Resumo:

Due to the evolution of technology, the speed and variety of collected data have increased in a surprising way, corroborating the creation of large and diverse datasets to be explored. These data usually need to pass through processes so knowledge can be extracted from them. For techniques that aim to perform predictions over the data, it is necessary that the data is properly annotated (or labeled). Labeling data samples can be challenging, expensive, and time-consuming. The lack of sufficient labeled data can make the application of supervised machine learning models impractical. Semi-supervised learning models have been extensively researched over the past decades. In the context of supervised learning, Gaussian Processes (GPs) offer the advantage of quantifying forecast uncertainties and providing significant modeling flexibility. However, like several Bayesian models, they cannot be directly applied to semi-supervised problems. This dissertation proposes a GP-based approach that allows its application on semi-supervised processes. The proposed method simultaneously trains an unsupervised GP latent variable model (GPLVM) and a supervised sparse GP model. The approach leverages both labeled and unlabeled data to create a more effective final classifier. Additionally, the approach incorporates a neural network to reproduce the latent variables learned by the GPLVM component, which enables its use with unseen data. The introduced approach is evaluated on public datasets and compared with other models from the literature.

Banca examinadora:

• Prof. Dr. César Lincoln Cavalcante Mattos (MDCC/UFC – Orientador)
• Prof. Dr. João Paulo do Vale Madeiro (MDCC/UFC)
• Prof. Dr. João Paulo Pordeus Gomes (UFC)