20 Mar 2024
14:00 Doctoral defense IC3 Auditorium
Theme
Building Convolutional Neural Networks from Markers
Student
Italos Estilon da Silva de Souza
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Deep Learning has revolutionized the area of machine learning, enabling and improving applications in various domains. Computer vision is one such domain that has been drastically transformed with deep learning techniques from large volumes of data -- annotated or not. However, the dependence on large volumes of data and, for some domains, the difficulty of acquiring, curating and/or annotating this data are limitations that compromise the cost and use of deep learning techniques. Another negative aspect is the lack of transparency in the model training process, which compromises their explainability and interpretability. To address these issues with deep learning models, specifically convolutional neural networks (CNNs), this doctoral thesis proposes human-guided learning methods for convolutional layer filters. The methodology that encompasses these methods is called Feature Learning from Image Markers (FLIM) and uses image markers indicated by humans in regions relevant to a task of interest - image classification or segmentation. FLIM is effective in creating simple image feature extractors (few weights and few layers) from markers (a very small set of data) that reach or surpass the performance of more complex models trained with more annotated data. FLIM can also be used to build models for image annotation in domains where annotation is laborious and susceptible to human error due to fatigue, such as remote sensing image segmentation. Thus, FLIM can be used to build a simple model that learns from a reduced set of annotated images and this model can annotate the rest of the images with pseudo labels which, in turn, can be used to train a more complex model ( deep) using the most confident labels. In this way, FLIM-based models can reduce human effort in annotating large volumes of data to train more complex models.
Examination Board
Headlines:
Alexandre Xavier Falcão | IC / UNICAMP |
João Paulo Papa | FC / UNESP |
Hélio Pedrini | IC / UNICAMP |
Esther Luna Colombini | IC / UNICAMP |
Thales Sehn Körting | INPE |
Substitutes:
Fabio Augusto Faria | ICT / UNIFESP |
Luiz Marcos Garcia Gonçalves | CT / UFRN |
Anderson de Rezende Rocha | IC / UNICAMP |
Sandra Eliza Fontes de Avila | IC / UNICAMP |