Diffusion-Based Sampling for Deep Active Learning

The remarkable performance of deep neural networks depends on the availability of massive labeled training data. To alleviate the load of data annotation with labels, deep active learning aims to sample a minimal set of training points to be labelled which yields maximal model accuracy. We propose an efficient sampling criterion to sample data for annotation, which automatically shifts from an exploration type of sampling to a class-decision-boundary refinement. Our criterion relies on a process of diffusing the existing label information over a graph constructed from the hidden representation of the data. This graph representation captures the intrinsic geometry of the approximated labeling function. We analyze our sampling criterion and its exploration - refinement transition in light of the eigen-spectrum of the diffusion operator. Additionally, we provide a comprehensive sample complexity analysis that captures the two phases of exploration and refinement. The diffusion-based sampling criterion is shown to be advantageous over state-of-the-art criteria for deep active learning on synthetic and real benchmark data.