Unsupervised Learning of Central Cases of Audio-Visual Events
Kento Nishibori, Tetsuya Matsumoto, Yoshinori Takeuchi, Hiroaki Kudo, Noboru Ohnishi · IEEJ Transactions on Electronics Information and Systems · 2009
In the real world, there are a lot of objects and it is impossible to make a system memorize all knowledge concerning the real world. Therefore, the system should autonomously learn knowledge relating to the environment. We propose a system that autonomously acquires concepts which are derived by statistical relation between audio-visual events. Firstly, the system determines correspondence between audio-visual events after extracting patterns from the external world, and accumulates them as cases. Secondly, it applies a canonical correlation analysis to the cases, and categorizes them by using K-means method. Finally, it identifies unknown image or sound, and associates the corresponding sound or image. As the result of experiments, the identification success rate of concepts is more than 83.2%. And the association success rate of concepts is more than 81.5%. Consequently, the effectiveness of this method was confirmed.