Parting with Illusions about Deep Active Learning

TL;DR

This paper critically evaluates deep active learning methods, revealing their limitations under realistic settings and proposing improved evaluation protocols to better assess their true effectiveness.

Contribution

It re-implements recent active learning approaches, assesses their performance in realistic scenarios, and introduces a more appropriate evaluation protocol.

Findings

Active learning methods are sensitive to training procedures.

Semi-supervised integration yields marginal improvements over random baseline.

Current evaluation schemes are inadequate for realistic assessment.

Abstract

Active learning aims to reduce the high labeling cost involved in training machine learning models on large datasets by efficiently labeling only the most informative samples. Recently, deep active learning has shown success on various tasks. However, the conventional evaluation scheme used for deep active learning is below par. Current methods disregard some apparent parallel work in the closely related fields. Active learning methods are quite sensitive w.r.t. changes in the training procedure like data augmentation. They improve by a large-margin when integrated with semi-supervised learning, but barely perform better than the random baseline. We re-implement various latest active learning approaches for image classification and evaluate them under more realistic settings. We further validate our findings for semantic segmentation. Based on our observations, we realistically assess the current state of the field and propose a more suitable evaluation protocol.