Exploring Uncertainty and Representativeness for Deep Active Learning

Albeit the considerable advances that are achieved by convolutional neural networks (CNN) in active learning, simultaneously investing multiple criteria such as uncertainty and representativeness when sampling informative images still remains a challenging issue. In this work, we develop and evaluate an active learning method called Exploring Uncertainty and Representativeness in Deep Active Learning (EURDAL), to iteratively boost the generalization capacity of the CNN backbone for image classification. To be specific, we introduce two adversarial image classifiers appended to the CNN backbone, and the uncertainty is expounded by the prediction discrepancy of the two adversarial classifiers. Besides, the ([Formula: see text])-tuplet loss, which was developed in deep metric learning, is imposed on the CNN backbone, to force CNN to extract discriminative features of unlabeled images. After that, the representativeness is defined by the distance proportion between the given example to its own centroid and the given example to all cluster centroids. With the integration of the learned uncertainty and representativeness, the most informative images will be selected while the noisy ones will be suppressed at each iteration of active supervision. Intensive image classification experiments are conducted on three benchmark datasets, and the encouraging results demonstrate the superiority and effectiveness of our proposed EURDAL network compared with some current competitive active learning methods.