Improving Data Labelling Efficacy with Deep Neural Network via Active Learning

This thesis revolutionizes Active Learning (AL) in Deep Learning, tackling the scarcity of labeled data. Introducing the Bayesian Estimate of Mean Proper Scores (BEMPS) and novel acquisition functions, it enhances AL across various domains. BEMPS is adapted for image classification with Monte-Carlo dropout and dialogue act classification, showing remarkable outcomes. The work also features a Deep AUC Maximization approach for Named Entity Recognition, leveraging AUC-Optimized Latent Sigmoid Embedding for precise uncertainty estimation. In its culmination, the thesis presents BESRA for multi-label text classification, extensively tested and proven on specialized medical datasets, marking a significant leap in AL's efficiency and applicability.