Active learning for object classification: from exploration to exploitation

摘要

Classifying large datasets without any a-priori information poses a problem in numerous tasks. Especially in industrial environments, we often encounter diverse measurement devices and sensors that produce huge amounts of data, but we still rely on a human expert to help give the data a meaningful interpretation. As the amount of data that must be manually classified plays a critical role, we need to reduce the number of learning episodes involving human interactions as much as possible. In addition for real world applications it is fundamental to converge in a stable manner to a solution that is close to the optimal solution. We present a new self-controlled exploration/exploitation strategy to select data points to be labeled by a domain expert where the potential of each data point is computed based on a combination of its representativeness and the uncertainty of the classifier. A new Prototype Based Active Learning (PBAC) algorithm for classification is introduced. We compare the results to other active learning approaches on several benchmark datasets.