Active Learning Thesis

Active learning is a subfield in machine learning in which a model will iteratively seek the most informative samples to include in its training dataset. Advances in this subfield has the potential to save millions of hours in data labelling costs for today’s data hungry deep learning models. This thesis aims to explore the use of active learning in object detection tasks; a largely unexplored application for active learning. Through our studies with the VOC2007 and MNIST datasets we have made three key contributions to the field. Firstly, we benchmarked the effectiveness of recent active learning techniques applied to object detection. Secondly, we provided an in-depth analysis on the type of data favored by each of these active learning techniques. Finally, we highlight the importance of re-initializing network weights after acquiring new data in an active learning scenario.