Large-Scale Visual Active Learning with Deep Probabilistic Ensembles
This work addresses the problem of efficient data annotation for deep learning practitioners, offering an incremental improvement in active learning methods for visual tasks.
The paper tackles the challenge of annotating data for training deep neural networks by introducing Deep Probabilistic Ensembles (DPEs), a scalable technique that approximates Bayesian Neural Networks for active learning. The result shows that DPEs require significantly less training data to achieve competitive performance on large-scale visual tasks like classification and semantic segmentation, steadily improving over baselines as annotation budget increases.
Annotating the right data for training deep neural networks is an important challenge. Active learning using uncertainty estimates from Bayesian Neural Networks (BNNs) could provide an effective solution to this. Despite being theoretically principled, BNNs require approximations to be applied to large-scale problems, where both performance and uncertainty estimation are crucial. In this paper, we introduce Deep Probabilistic Ensembles (DPEs), a scalable technique that uses a regularized ensemble to approximate a deep BNN. We conduct a series of large-scale visual active learning experiments to evaluate DPEs on classification with the CIFAR-10, CIFAR-100 and ImageNet datasets, and semantic segmentation with the BDD100k dataset. Our models require significantly less training data to achieve competitive performances, and steadily improve upon strong active learning baselines as the annotation budget is increased.