Self-learning with confidence bands

Since manual labelling of huge data sets is costly and time-consuming, we propose a framework for iterative confidence-based self-learning of classifiers which autonomously extends its knowledge gained based on a small amount of initial, manually labelled training samples towards increasingly different representatives of the regarded pattern classes. During each iteration, the labels of newly selected samples are determined and the samples are added to the training set, subsequently. The crucial step is the selection of the samples which are used for re-training the classifier during each iteration. For this purpose, we rely as a new approach on the concept of confidence bands known from the field of statistics.