Enhanced Active Learning Through Exclusion of Semi-Informative Sets

초록

In this study, we explore methods for minimizing labeling costs, which is a critical issue in the active learning process. We focus on efficiently selecting the most essential datasets from an entire pool of unlabeled datasets. Alongside previous methodologies that sample datasets based on model uncertainty and feature map diversity, we propose a novel approach aimed at budget optimization. Notably, we introduce the concept of semi-informative sets, which allows for the effective exclusion of data that contribute relatively little to learning. This approach aims to maximize budget efficiency, leading to two main strategies: The first strategy involves refining the prediction probabilities of the model by applying a label-smoothing technique to improve the loss function that is directly associated with the prediction accuracy. The second strategy centers on the effective selection of high-confidence data through a more precise threshold setting. To achieve this, we introduce a learnable threshold that is adjusted based on the reliability of the already learned data. These approaches present a novel direction for active learning that considers budget constraints and potentially contributes to the development of practical active learning systems for real-world applications.

키워드