Causality detected by transfer entropy leads acquisition of joint attention

Abstract

Joint attention, i.e., the behavior of looking at the same object another person is looking at, plays an important role in both human communication and human-robot communication. Previous synthetic studies have focused on modeling the developmental process of joint attention and have proposed learning methods without any explicit instructions for joint attention. The causal structure between a perception variable (the caregiver's face directions or individual objects) and an action variable (gaze shift to the caregiver's face or object locations) is given in advance to learn joint attention. However, such a structure is expected to be found by the robot through the interaction experiences. This paper investigates how the transfer entropy, that is an information theoretic measure, can be used to quantify the causality inherent in the face-to-face interaction. In the computer simulation of human-robot interaction, we examined which pair of perceptions and actions are selected as the causal pair and showed that the selected pairs can be used to learn a sensorimotor map for achieving joint attention.