In networks with a conventional learning method based on correlation matrixes,the leaning stages are separated from the remembering stages. But it is considered that they are not separated in the brain. Therefore, the studies of networks' learning methods close real brain's networks are advanced. In this study, the experiments of the networks with the learning methods proposed by Asakawa et al. are made. This network distinguishes an unknown pattern from the known patterns, and if it is the unknown pattern, the network performs supplementary learning.When a known pattern is input into this network, it searches around the input pattern. In contrast, when an unknown pattern is input, the network shows chaotic movement between this pattern and learning pattern. Making use of the property, the network distinguishes the unknown patterns from the known patterns. Furthermore, input pattern is given continuously, and outside input term is added. This network can then learn a correct pattern even if an unknown pattern with noise is input into it.
In this study, localized supplementary learning method is compared with a conventional Hopfield neural net's learning method. and examined. Through computer simulations, the effectiveness of supplementary learning method is shown.
競合作用