把持運動において対象物の画像情報の変更が手形状計算へ与える影響

抄録

From the viewpoint of computational theory, it has been postulated herein that certain properties of objects are extracted from their visual images for visuo-motor transformation in human grasping movements. To examine the computational feasibility of this hypothesis, a five-layer neural network model has been constructed for transforming the visual images of objects into the prehensile hand shape. After sufficient learning, the property information corresponding to the size and shape of objects was extracted in the third layer. The values of the third-layer and output-layer units changed widely for shadingless images and only slightly for blurred images, in comparison with the original images. Essentially the same results were obtained from a psychophysical experiment conducted on human grasping movements with the same images (original, shadingless and blurred) as those used in the neural network study, thus supporting the plausibility of the present computational hypothesis.