Learning action sequences using recurrent neural network with multiresolutional channel
Satoru Morita · Systems and Computers in Japan · 2001
Abstract The author proposes a recurrent neural network structure that enables associations between resolution, field of view, and actions in order to connect actions with the surrounding view for executing multiresolutional vision control. In general, when a neural network is used, there are cases when some patterns cannot be learned. The current study describes obtaining surrounding views necessary for capturing actions by using a hereditary algorithm in order to control multiresolutional vision. The article demonstrates how in practice actions of a pan–tilt camera can be captured using multiresolutional vision control. © 2001 Scripta Technica, Syst Comp Jpn, 32(13): 31–40, 2001