Advances in Deep Active Learning and Synergies With Semi-Supervision
MCML Authors
Sandra Gilhuber (née Obermeier)
Abstract
Sandra Gilhuber (née Obermeier)
Abstract
This thesis tackles the challenge of costly and limited annotated data for training deep learning models by advancing methods in deep active learning and its combination with semi-supervised learning. The work introduces efficient strategies to select informative and diverse samples without expensive computations, achieving superior performance in both tabular and image classification tasks. It further develops a novel approach for active learning on graphs, leveraging diffusion-based heuristics for node selection, and proposes refined pseudo-labeling techniques to mitigate confirmation bias in semi-supervised settings. Altogether, the contributions demonstrate how integrating active and semi-supervised learning can reduce labeling costs, improve efficiency, and address real-world data challenges. (Shortened.)
BibTeXKey: Gil25