Learning Correspondences Between Visual Features and Functional Features

We have implemented a visual learning system MIRACLE-IV, which is capable of obtaining an internal structure of an object from a series of silhouette images with no initial explicit models about the object[1, 2,3]. The images are derived from only one object, but the forms of the object are varied. The system is composed of two sub-systems: a model-acquisition part (the modeler) and an image-processing strategy part (the strategist). On the assumption that the object consists of hinges, slides and solids, the modeler learns the number of them in the object and the relationship between them. The strategist binds the functional features as hinges or slides with visual features in the actual image data. The image-processing sequence for the extraction of the visual feature is not given previously but is learned automatically through trial and error. In our research, mutual references between pattern information and symbol description play essensial roles for learning. This paper describes how MIRACLE-IV learns correspondence between fuctional features and visual features.