Abstract:

Recent years have witnessed significant breakthroughs in deep learning, but the efficacy of these techniques is contingent on clean data. Unfortunately, in numerous practical applications, label ambiguity can stem from sources such as labeling inconsistencies, ill-posedness, sampling bias, or errors during data collection. Ambiguous labels can hamper the performance of machine learning models, leading to inaccurate classifications or predictions. In this talk, Kshitij will present his work, which focuses on developing a novel generation of machine learning algorithms to manage ambiguous labels in text, imaging systems, and environmental modeling tasks.

Location:

Microsoft Teams