Active Learning of Pattern Classification Based on PEDCC-Loss

Abstract

Deep learning classifiers require a large number of labeled samples to train the model. Active learning reduces the dependence of classification model on labeled samples by gradually selecting high-quality samples for iterative training. In this article, an active learning method for pattern classification is proposed, which can use fewer labeled samples to achieve higher classification accuracy. The algorithm uses PEDCC-Loss to train the network model according to the predefined evenly-distributed class centroids (PEDCC) of the classification task, maximizing the distance between classes. Then, the maximum value of the cosine distance is selected to measure the sample uncertainty according to the PEDCC output characteristics. In the active learning algorithm iteration, samples with high uncertainty are selected for manual labeling, and samples with low uncertainty are given pseudo labels. At the same time, manual label samples and pseudo label samples are used to train the classification network to continuously optimize the classification boundary. Compared with CRL, LLAL and other methods, the experimental results on CIFAR100, CIFAR10, SVHN, and Animals-10 data sets show the effectiveness of this algorithm.

Keywords

• Active learning
• predefined class centroids
• PEDCC-loss
• pattern classification