Multimodal joint attention through cross facilitative learning based on μX principle

Simultaneous learning of multiple functions is one of the fundamental issues not only to design intelligent robots but also to understand human’s cognitive developmental process since we, human, do so in our daily lives but we do not know how to do. Drawing an analogy to the well-known bias in child language development, we propose the mutual exclusivity selection principle (μX principle) for learning multi-modal mappings: selecting more mutually exclusive output leads experiences to make underdeveloped complementary mappings more disambiguated. The μX principle is applied to multi-modal joint attention with utterances for lexicon acquisition, and synthetically modeled in both intraand inter-module levels of output. Through the series of computer simulations, the effects of the μX principle on the mutual facilitation in learning multi-functions and robustness against errors in segmentation of observation are analyzed. Finally, the correspondence of the synthesized development to infant’s one is argued based on the simulation with careful behavior by a caregiver.