Nearest Neighbor Classifier Embedded Network for Active Learning

Deep neural networks (DNNs) have been widely applied to active learning. Despite of its effectiveness, the generalization ability discriminative classifier (the softmax classifier) is questionable when there a significant distribution bias between labeled set and unlabeled set. In this paper, we attempt replace in deep network with nearest neighbor classifier, considering progressive within unknown sub-space. Our proposed learning approach, termed Neighbor Classifier Embedded (NCE-Net), targets at reducing risk over-estimating samples while improving opportunity query informative samples. NCE-Net conceptually simple but surprisingly powerful, as justified from perspective subset information, which defines metric quantify model Experimental results show that, selection based on rejection or confusion confidence, improves state-of-the-arts image classification object detection tasks margins.