Iker Beñaran Muñoz: "On the Study of Crowdsourced Labelled Data and Annotators: Beyond Noisy Labels“.

Zuzendariak_Directores: Aritz Pérez Martínez/ Jerónimo Hernández González.

2023_05_12, 11:00  Sala Ada Lovelace aretoa.

Abstract:

"Crowdsourcing is nowadays used extensively for gathering labels for supervised classification due to its availability and cheapness. For this reason, many  researchers have proposed diverse approaches to learning from crowds. In this thesis, we provide novel techniques for improving the efficiency of learning from crowds. The aim is to profit from the available information as much as possible, either from the descriptive feature of the instances or from the doubts of the annotators. For that, in this thesis we propose a label aggregation method, a new flexible crowd labelling scheme with two associated aggregation techniques, and classifier learning methods that make use of the novel labelling scheme. The techniques developed in this thesis show that competitive results can be reached involving reduced costs, enhancing efficiency in learning from crowds and opening the path for promising research lines."