Inductive conformal predictor for convolutional neural networks: Applications to active learning for image classification

Abstract

Conformal prediction uses the degree of strangeness (nonconformity) of data instances to determine the confidence values of new predictions. We propose an inductive conformal predictor for convolutional neural networks (CNNs), referring to it as ICP-CNN, which uses a novel nonconformity measure that produces reliable confidence values. Furthermore, ICP-CNN is used to improve classification performance through active learning, selecting instances from an unlabeled pool based on the evaluation of three criteria: informativeness, diversity, and information density. Distance metric learning is employed to measure diversity, using a similarity measure that adapts to the database being used. Moreover, information density is considered to filter outliers. Experiments conducted on face and object recognition databases demonstrate that ICP-CNN improves the classification performance of CNNs, outperforming previously proposed active learning techniques, while producing reliable confidence values.