Visual Active Learning guided by Feature Projections

Abstract

Machine Learning models can be very effective when the number of labeled training samples is high. In some areas, such as Medicine and Biology, large labeled training sets are very difficult to be obtained, given that the manual annotation of such datasets is time-consuming and requires specialists. In general, data annotation based on the visual inspection (supervision) of each training sample (object) is a laborious process, especially when the number of required samples is large. In order to mitigate user effort in data annotation, many studies have adopted active learning techniques in which an apprentice classifier suggests informative samples for user supervision. It is expected that the classifier, retrained with those additional annotated samples, performs better in a next iteration. However, finding a small and effective set of informative samples for user supervision may be difficult since it also depends on the feature space. As a result, the number of active learning iterations (number of supervised samples) and user involvement can still be high. With similar aim, we have presented interactive data annotation techniques guided by feature projection as part of the Master's thesis of the candidate. In these works, the user can visualize supervised (small set) and unsupervised (large set) training samples on a 2D projection space and propagate labels to the unsupervised ones, with and without assistance of a pattern classifier, but with no further object supervision. In this Doctoral proposal, we intend to extend our previous work by investigating and developing the combination of visual analytics, feature learning, and active learning techniques for the design of more effective image classification systems with minimum user effort in data annotation. The proposal also includes a period abroad at the University of Utrecht under the supervision of Prof. Alexandru Telea --- a recognized expert in Visual Analytics.