Active one-shot learning for personalized human affect estimation

Abstract

Building models that can classify human affect leads to the challenge of learning on data that is complex in features and limited in size and labels. How can these models balance being general and personalized, capturing both the commonalities and the individual quirks of people? While previous research has explored the intersection of deep learning, active learning, and one-shot learning to craft models that are semi-supervised and data-efficient, these methods have not yet been examined in the context of personalized affective computing. This study presents a novel active one-shot learning model for personalized estimation of human affect, in particular, detection of pain from facial expressions. The model demonstrates the ability to learn an active learner that achieves high accuracy, learns to become data efficient, and introduces model personalization to match or outperform fully supervised and population-level models.