Beam Search Optimized Batch Bayesian Active Learning

Active Learning is an essential method for label-efficient deep learning. As a Bayesian active learning method, by Disagreement (BALD) successfully selects the most representative samples maximizing mutual information between model prediction and parameters. However, when applied to batch acquisition mode, like construction with greedy search, BALD suffers from poor performance, especially noises of near-duplicate data. To address this shortcoming, we propose diverse beam search optimized which explores graph every expanding highest-scored predetermined number. avoid near duplicate branches (very similar beams generated same root samples), undesirable lacking representations in feature space, design self-adapted constraint within candidate beams. The proposed able acquire data that can better represent distribution unlabeled pool, at time, be significantly different existing We observe achieves higher performance than baseline methods on three benchmark datasets.