Keywords: machine learning, active learning, uncertainty, hypothesis perturbation, disagreement region
Abstract: Active learning strategy to query samples closest to the decision boundary can be an effective strategy for sampling the most uncertain and thus informative samples. This strategy is valid only when the sample's "closeness" to the decision boundary can be estimated. As a measure for evaluating closeness to a given decision boundary of a given sample, this paper considers the least probable disagreement region (LPDR) which is a measure of the smallest perturbation on the decision boundary leading to altered prediction of the sample. Experimental results show that the proposed LPDR-based active learning algorithm consistently outperforms other high performing active learning algorithms and leads to state-of-the-art performance on various datasets and deep networks.
One-sentence Summary: This paper defines a measure of sample's closeness to the decision boundary based on the disagree metric between hypotheses in active learning.
12 Replies
Loading