Our visual sensor consists of many retinal neurons, each of which covers only a local receptive field. In order to recognize the object location as an analog value, we have to integrate visual signals from many retinal neurons. In our real world, objects usually move smoothly according to the equations of motion. If an output of a neural network which receives inputs from retinal neurons, changes smoothly along time, the output represents the object location as an analog value. Then the learning method has been proposed, which makes the output of the neural network become smooth, in other words, makes the second time derivative of the output become 0.
抄録全体を表示