Dataset refinement for convolutional neural networks via active learning

Abstract

Convolutional neural networks (CNNs) have shown significant advantages in computer vision fields. For the optimizations of CNNs, most research works focus on feature extraction, which creates deeper structures and better activations, but the optimizations on dataset is rarely discussed. Due to the boom of data, most CNNs suffer from serious problems of dataset redundancy and following high computational burden. To this end, this paper brings in an informativeness ranking thought and proposes a new methodology for dataset refinement based on Active Learning. Extensive experiments prove its effectiveness to achieve a higher classification accuracy for CNNs at a less training cost. Moreover, for classification problems with a large number of class, this paper further proposes Entropy Ranking, a new Active Learning method, to enhance the optimization ability.