Diverse Bayesian Active Learning with Simulated Annealing for Probabilistic Sampling

We propose a simple yet highly efficient and robust active learning (AL) framework for image classification. Most of the existing AL strategies are either not scalable with increasing acquisition batch sizes or not robust to noise. They select samples greedily without considering the acquisition state of previous iteration. Further, very little focus has been given to the selection of the initial seed set for active learning. In this work, we propose a new framework that combines simulated annealing within AL to select those samples which improve their acquisition cost in the previous iteration. A convex combination of a diversity measure and an uncertainty measure is used as the acquisition cost. The diversity measure ensures consistent prediction of samples lying farthest from the decision boundaries and, eventually, an unbiased estimation of uncertainty. We demonstrate the efficiency and robustness of our proposed framework over the current state of the art AL strategies using Bayesian CNNs.