Direct Acquisition Optimization for Low-Budget Active Learning

Active Learning (AL) has gained prominence in integrating data-intensive machine learning Many active learning algorithms have emerged over the (ML) models into domains with limited labeled past decades, with early seminal contributions from (Lewis, data. However, its effectiveness diminishes significantly 1995; Tong & Koller, 2001; Roy & McCallum, 2001), and a when the labeling budget is low. In shift that focuses more on deep active learning - a branch this paper, we first empirically observe the performance of AL that targets more towards DL models in more recent degradation of existing AL algorithms years (Huang, 2021). Depending on the optimization in the low-budget settings, and then introduce objective, AL algorithms can be classified into two categories. Direct Acquisition Optimization (DAO), a novel The first category includes heuristic objectives that AL algorithm that optimizes sample selections are not exactly the same as the evaluation metric, i.e. error based on expected true loss reduction.