摘要
We propose a learning architecture for integrating multi-modal information e.g., vision, audio information. In recent years, artificial intelligence (AI) is making major progress in key tasks like a language, vision, voice recognition tasks. Most studies focus on how AI could achieve human-like abilities. Especially, in human-robot interaction research field, some researchers attempt to make robots talk with human in daily life. The key challenges for making robots talk naturally in conversation are to need to consider multi-modal non-verbal information same as human, and to learn with small amount of labeled multi-modal data. Previous multi-modal learning needs a large amount of labeled data while labeled multi-modal data are shortage and difficult to be collected. In this research, we address these challenges by integrating single-modal classifiers which trained each modal information respectively. Our architecture utilized knowledge by using bi-directional associative memory. Furthermore, we conducted the conversation experiment for collecting multi-modal non-verbal information. We verify our approach by comparing accuracies between our system and conventional system which trained multi-modal information.
