音声模倣と語彙獲得の共発達のための主観的整合機構に基づく対応学習

抄録

This paper presents a learning mechanism that enables co-development of vocal imitation and lexicon acquisition by integrating multimodal information based on subjective consistency. Infant-caregiver interaction is often assumed in modeling infant development. A caregiver is basically assumed to react to an infant in a teaching manner, for example, imitating the learner's voice and labelling an object that it is looking at. However, such tendency is not always expected. Subjective consistency is introduced to judge whether to believe the observed experiences (external input) as a reliable signal for learning. The learning mechanism estimates the outputs of one layer by combining that of other layers and an external input. Based on the proposed mechanism, a simulated infant robot learns associations from its caregiver's phonemes, its own phonemes and objects to co-develop the vocal imitation and lexicon acquisition. The result of computer simulation indicates that the proposed mechanism realized mappings for vocal imitation and lexicon acquisition even when a caregiver does not always react to a learner in a teaching manner.