This paper proposes new incremental learning methods of generalized radial basis function (GRBF). We define a magnitude of interference as the norm of the distance between the output functions before learning and the one after learning. During the learning phase, the GRBF modifies its parameters so as to minimize an objective function described by the sum of an error function and the magnitude of interference.
It is supposed that the human recognizes grasped objects using visual information and somato-sensory information. It should be noticed that the relation between these kinds of sensory information about an object is many-to-many relation and the human must recognize the grasped object from such kinds of information. In our previous work, we have proposed a neural network model that makes an internal representation of an object by integrating visual information and somato-sensory information about the grasped object. In this paper, we confirm that our proposed neural network model can learn the many-to-many relation between these kinds of information using numerical simulation.