Self- And Weakly- Supervised Deep Learning Methods With Applications In Biometric And Biomedical Data

This dissertation introduces novel deep learning methodologies for effectively leveraging weakly-labeled biomedical data and uncurated/unlabeled biometric data. The thesis is divided into three major parts. In the first part, we present a classifier that combines 2D and 3D classifiers that are trained with weak supervision using volume-wise labeled CT lung images. The main contribution of the thesis is a new representation learning method, extending the contrastive learning framework with the variational approach. In the second part of the thesis, we present a semisupervised approach using the variational contrastive design, applied to learning face attributes from web-collected face images. This technique, called VCL-PL, is specifically designed to counter the inherent noise found in the collected images. Through various experimental setups, the method demonstrates an enhancement in accuracy over supervised or state-of-the-art self-supervised methods. The last part of the dissertation develops a robust self-supervised learning model, VCL, that combines variational contrastive learning with beta-divergence. This model exhibits better performance than state-of-the-art models when used with unlabeled, uncurated, and noisy datasets. Through the development of these methodological advancements and the introduction of novel datasets, this dissertation contributes to learning from weakly-labeled data in the medical domain and introduces the variational contrastive learning approach that better handles noisy data and low data regimes, in the biometric domain.