Active Learning with Aggregated Uncertainties from Image Augmentations

Active learning and data augmentation are both standard techniques for dealing with a lack of annotated data in the field of machine learning. While active learning aims to select the most informative data sample for annotation from a pool of unlabeled data, data augmentation enhances the data set’s volume and variety, introducing modified versions of existing data. We propose a method that combines both approaches and exploits their benefits beyond mere data quantity by taking into account the relationship of original image and augmentation tuples from the perspective of the underlying machine learning model. Namely, we explore the distribution of uncertainties within these tuples and their effect on model performance. Our research shows that with equal annotation effort aggregated uncertainties across image augmentations yield improved results compared to a baseline without augmentations, however certain configurations can be detrimental for the performance of the resulting model.

VIVO