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

作者：

Highlights：

• Inductive conformal predictors using nonconformity measures designed for convolutional neural networks produce reliable confidence values

• The combination of informativeness, diversity, and information density in a single query function improves the performance of active learning

• Distance metric learning produces similarity measures that adapt to the databases being used, improving the performance of query functions for active learning

• Dimensionality reduction through principal component analysis significantly reduces the computational load of distance metric learning

摘要

•Inductive conformal predictors using nonconformity measures designed for convolutional neural networks produce reliable confidence values•The combination of informativeness, diversity, and information density in a single query function improves the performance of active learning•Distance metric learning produces similarity measures that adapt to the databases being used, improving the performance of query functions for active learning•Dimensionality reduction through principal component analysis significantly reduces the computational load of distance metric learning

论文关键词：Conformal prediction,Convolutional neural networks,Active learning,Distance metric learning,Image classification

论文评审过程：Received 17 July 2018, Revised 14 January 2019, Accepted 25 January 2019, Available online 27 January 2019, Version of Record 1 February 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.01.035