Reinforcement Learning Based Online Active Learning for Human Activity Recognition

Abstract

Online active learning (OAL), i.e., asking a user in a targeted and parsimonious way to provide annotation for activities they are currently engaged in, has been established as a meaningful way for bootstrapping human activity recognition (HAR) systems for real-world deployments. In this paper we extend on the idea of optimizing budgets of user-provided annotations by introducing a reinforcement learning based OAL approach. Our method decides on which data sample a user shall provide a label for using a continuosly updated base classifier and a reward function that takes into account the classifier's confidence in form of its a-posteriori probability. We evaluate our approach on seven benchmark datasets and demonstrate recognition capabilities of the resulting classifiers that are superior to the state-of-the-art and reach the performance of fully supervised baseline systems for half the datasets. The presented approach has the potential to push the boundaries for real-world deployments of HAR systems.