Unsupervised active learning with loss prediction

Active learning is an effective technique to reduce the cost of labeling data by selecting the most beneficial samples. Most existing active learning methods use linear models to select the most representative points to approximate other points. However, they only select samples from the perspective of informativeness or representativeness and cannot model the nonlinearity of data well. In this paper, we propose a novel unsupervised active learning method with a loss prediction module, called UALL. Specifically, UALL uses a deep neural network to model the nonlinearity of data and considers simultaneously the representativeness, informativeness, and diversity, three essential criteria in active learning. Furthermore, we introduce an autoencoder and a loss prediction module to evaluate the representativeness and informativeness and combine K-means and simple calculations to measure the diversity. We compare with the state-of-the-art on eight publicly available datasets from different fields, and the experimental results demonstrate the effectiveness of our method.

Automated summary

This paper proposes a novel unsupervised active learning method that models the nonlinearity of data using a deep neural network and considers essential criteria such as representativeness, informativeness, and diversity. Experimental results demonstrate the effectiveness of the method.