Discriminator-based active learning with feature decomposition

Abstract

When training a model, the most important things are the amount and quality of data. However, it is not always possible to obtain a large amount of data in real life. In addition, data is often not well labeled. For example, in the medical field, doctors determine the part of a patient's disease through meetings with various doctors to determine the part of the disease. Ordinary people cannot easily label medical images. In addition, when analyzing the enemy, tank and weapon information of the enemy is confidential, so it is difficult to label what kind of weapon it is and expert help is essential. In this case, labeling is very expensive and active learning, which selects data that is helpful for learning, is a very important field. In general, active learning research has been conducted focusing on uncertainty within the range that humans can predict. However, there may be important information for labeling, so additional research is needed using a network-based discriminator for active learning. In this paper, we have developed a structure that extracts the necessary information for labeling using mutual information and cycle consistency without damaging the original information. We experimentally showed that mutual information and image reconstruction loss converged well, and as a result, the accuracy was higher in CIFAR 10 compared to the existing algorithm(VAAL). The accuracy was also shown to quickly increase in the KAIST-Army dataset. If this algorithm is applied to the expensively labeled medical or defense fields, it can quickly find data that is informative for classification and label it, enabling models with low accuracy that could not be used to solve actual problems quickly to be used for solving actual problems in the medical or defense fields where data labeling is expensive.