連続時間ダイナミクスをもつ連想記憶によるマルチモーダル情報処理

抄録

It is difficult to realize the context dependency of cognition in real world environment using conventional pattern recognition technique. We have proposed memory model PATON that realizes such cognition based on attention. To apply the PATON to real environment, we introduce a continuous-time dynamics, a mechanism of event driven attention generator, and an autoassociator network with constraints for encoding of external input in various modality. Then it is shown that the ability of multi-modal information processing of PATON is useful for continuous spoken word recognition task with visual input.