Multi-Classifier Adversarial Optimization for Active Learning

Active learning (AL) aims to find a better trade-off between labeling costs and model performance by consciously selecting more informative samples label. Recently, adversarial approaches have emerged as effective solutions. Most of them leverage generative networks align feature distributions labeled unlabeled data, upon which discriminators are trained distinguish them. However, these methods fail consider the relationship decision boundaries, their training processes often complex unstable. To this end, paper proposes novel AL method, namely multi-classifier optimization for active (MAOAL). MAOAL employs task-specific boundaries data alignment while most fulfill this, we introduce classifier class confusion (C3) metric, represents discrepancy inter-class correlation outputs. Without any additional hyper-parameters, C3 metric further reduces negative impacts ambiguous in process distribution sample selection. More concretely, network is adversarially adding two auxiliary classifiers, reducing bias minimizing loss tighter highlighting hard maximizing loss. Finally, with highest selected Extensive experiments demonstrate superiority our approach over state-of-the-art terms image classification object detection.