Modèle crédibiliste pour l’échantillonnage en apprentissage actif


In machine learning, training a classifier on large dataset requires an important amount of labels which is expensive in terms of human resources and money. A possible solution to this problem is to use crowdsourcing in order to label the data. Although, non-expert people do not always have the knowledge to do their work correctly, leading to introducing errors in labeling. Active learning offers a solution to the labeling cost by making the classifier choose the data it wants to label in order to reach good performance with fewer labels. By combining active learning and belief functions, it becomes possible to model the errors and uncertainty in labels. We propose a new sampling method implying belief entropies.