Active learning aims to select samples to be annotated that yield the largest\nperformance improvement for the learning algorithm. Many methods approach this\nproblem by measuring the informativeness of samples and do this based on the\ncertainty of the network predictions for samples. However, it is well-known\nthat neural networks are overly confident about their prediction and are\ntherefore an untrustworthy source to assess sample informativeness. In this\npaper, we propose a new informativeness-based active learning method. Our\nmeasure is derived from the learning dynamics of a neural network. More\nprecisely we track the label assignment of the unlabeled data pool during the\ntraining of the algorithm. We capture the learning dynamics with a metric\ncalled label-dispersion, which is low when the network consistently assigns the\nsame label to the sample during the training of the network and high when the\nassigned label changes frequently. We show that label-dispersion is a promising\npredictor of the uncertainty of the network, and show on two benchmark datasets\nthat an active learning algorithm based on label-dispersion obtains excellent\nresults.\n