Hypothesis Perturbation for Active Learning

This paper introduces a computationally efficient Query-by-Committee (QBC) algorithm specifically designed for deep active learning. The algorithm leverages the concept of hypothesis perturbation (HP) to construct the committee. The conventional QBC algorithms often incur high computational costs due to the independent training required for each committee member. In contrast, the HP constructs the committee by strategically sampling hypotheses around a given hypothesis, and efficiently identifies data points located near the decision boundary of the current hypothesis. To quantify uncertainty, the algorithm leverages a novel metric termed disagreement in hypothesis perturbation (DHP). DHP quantifies the disagreement in predictions between the given hypothesis and its perturbed hypotheses. This metric effectively identifies data points with high uncertainty, making them ideal candidates for active learning. The effectiveness of the proposed DHP-based active learning algorithm is empirically validated through extensive experimentation. The results demonstrate that the algorithm consistently achieves superior performance compared to other established algorithms across various datasets and deep network architectures considered in the study.