ActiveConfusion: Efficient Cold-Start Active Learning Using Pretext Task Confusion

Azeez Idris, Wallapak Tavanapong

Abstract: Active Learning (AL) reduces manual annotation costs by iteratively selecting a small, but most informative subset of an unlabeled dataset for labeling by domain experts. The quality of the initial subset can significantly impact the entire AL process. The problem of selecting the initial subset is known as the cold-start problem. Recent AL methods, e.g., PT4AL, have used self-supervised learning (SSL) to address the cold-start problem. However, SSL can be computationally expensive to train. In this paper, we present ActiveConfusion, a simple and time-efficient strategy for addressing the cold-start problem in vision Active Learning. The proposed approach requires only a single epoch of SSL pretext task training to derive a confusion score to rank the unlabeled images. Performance evaluation of ActiveConfusion versus the cold start method in PT4AL on five natural and medical image datasets show promising results while significantly reducing time.