Active learning methods with deep gaussian processes

Abstract

Active learning is an effective method to reduce the learning time, space and economic costs in the whole training procedure. It aims to select more informative points from the unlabeled data pool, label them and add them into the training set, which helps to improve the performance of learning models. Learning models and active learning strategies are two essential elements in the framework of active learning. Probabilistic models such as Gaussian processes are often used as learning models for active learning, which have achieved promising results attributed to their predictive uncertainty. In order to well model complex data and characterize uncertainty, we employ deep Gaussian processes (DGPs) as learning models, based on which active learning strategies are made. Specifically, we design appropriate active learning strategies based on DGPs for solving binary and multi-class classification tasks, respectively. The experiments on educational and non-educational text classification and handwritten digit recognition demonstrate the effectiveness of the proposed active learning methods.