Interactive Visualization based Active Learning
Mohammadreza Babaee, Stefanos Tsoukalas, Gerhard Rigoll, Mihai P. Datcu · elib (German Aerospace Center) · 2014
Active learning aims to label the most informative data points in order to minimize the cost of labeling [1]. In this work, we introduce a human based approach, namely First Certain Wrong Labeled (FCWL) to select points for labeling. It is based on a ranked list of predictions ordered by confidence, from which the user selects the highest ranking incorrect prediction. The experimental results show the improvement in performance of this method compared to others.