Combining Human and AI Strengths in Object Counting under Information Asymmetry

Abstract

With the recent development of artificial intelligence (AI), hybrid human-AI teams have gained more attention and have been employed to solve all kinds of problems. However, existing research tends to focus on the setting where the same task is given to humans and AI. This work investigates a scenario where different agents have access to different types of information regarding the same underlying problem. We propose a probabilistic framework that combines the predictions of humans and AI based on the quality of the information given by both agents. We apply this framework to a regression task in which humans and AI are given different views of a jar and aim to estimate the number of objects in it. We demonstrate that our model can outperform methods that ignore information asymmetry. Furthermore, we show that complementarity can be achieved, i.e., combining human and AI predictions leads to better performance than relying on humans or AI alone. This framework can be adapted to solve other problems in which different sources of information from multiple agents are present.