| 摘要 |
Starting from a device with no hidden neuron, arbitrary samples are taken from a learning sample set and are presented as objects to be classified. On each occasion, if the response is not correct, a hidden neuron (Hi) is introduced together with a connection to the output neuron (Oj) of the class of the sample, whereas if the response is correct no neuron is added. During this phase of introducing hidden neurons, the neurons are divided into groups by investigating, for each neuron introduced, whether it falls inside an existing group, in which case it is incorporated into it, otherwise a new group is created around it. Membership of a group is defined on the basis of the distance from the "creator" neuron. Applications to character recognition systems. <IMAGE> |
