Deep Active Ensemble Sampling for Image Classification

Conventional active learning (AL) frameworks aim to reduce the cost of data annotation by actively requesting labeling for most informative points. However, introducing AL hungry deep algorithms has been a challenge. Some proposed approaches include uncertainty-based techniques, geometric methods, implicit combination and approaches, more recently, based on semi/self supervised techniques. In this paper, we address two specific problems in area. The first is need efficient exploitation/exploration trade-off sample selection AL. For this, present an innovative integration recent progress both enable exploration/exploitation strategy. To end, build computationally approximate Thompson sampling with key changes as posterior estimator uncertainty representation. Our framework provides advantages: (1) accurate estimation, (2) tune-able between computational overhead higher accuracy. second problem improved training protocols use ideas from propose general approach that independent technique being used. Taken these together, our shows significant improvement over state-of-the-art, results are comparable performance supervised-learning under same setting. We show empirical framework, comparative state-of-the-art four datasets, namely, MNIST, CIFAR10, CIFAR100 ImageNet establish new baseline different settings.