When Deep Learners Change Their Mind: Learning Dynamics for Active Learning

Active learning aims to select samples be annotated that yield the largest performance improvement for algorithm. Many methods approach this problem by measuring informativeness of and do based on certainty network predictions samples. However, it is well-known neural networks are overly confident about their prediction therefore an untrustworthy source assess sample informativeness. In paper, we propose a new informativeness-based active method. Our measure derived from dynamics network. More precisely track label assignment unlabeled data pool during training We capture with metric called label-dispersion, which low when consistently assigns same high assigned changes frequently. show label-dispersion promising predictor uncertainty network, two benchmark datasets algorithm obtains excellent results.