Fig. 7.

Active learning results on the CIFAR-10 dataset. The accuracy improvement of our approach over previous state-of-the-art methods is most significant when training with a limited number of samples (2k and 3k out of total 50k images), similar to the annotation budget for EM-R50 (≈ 5%). Mean and standard deviation are estimated from 5 runs. We also show that the accuracy saturates after ten iterations of query suggestion (Fig. S-4 in the supplementary material).