Large-Scale Visual Active Learning with Deep Probabilistic Ensembles

TL;DR

This paper introduces Deep Probabilistic Ensembles (DPEs), a scalable method for active learning in large-scale visual tasks that achieves competitive performance with less labeled data by effectively estimating uncertainty.

Contribution

The paper proposes DPEs, a scalable ensemble-based approach that approximates deep Bayesian Neural Networks for large-scale active learning tasks.

Findings

DPEs require less training data to reach competitive accuracy.

DPEs outperform strong active learning baselines as annotation budgets grow.

Effective uncertainty estimation improves active learning efficiency.

Abstract

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.