Towards Visual Explainable Active Learning for Zero-Shot Classification

Abstract

Fig. 1. Semantic navigator is a mixed-initiative visual analytics system for zero-shot classification. (a) The machine asks contrastive questions to guide analysts to come up with new attributes. (b) The semantic map explains the machine's status and presents the label recommendations (striped contours). Analysts select partial classes as positive (solid red contours) or negative (solid blue contours) to adjust the label recommendations ((c) and (d)). (e) The line chart monitors the training accuracy for seen classes and testing accuracy for unseen classes. (f) The class-attribute matrix is built interactively via collaboration between analysts and the machine.