Incremental learning is a learning method used to chaotic neural network. The existing learning method to neural network is correlation learning. The neural network can memorize more patterns with incremental learning against correlation learning. Whereas it is known that the neural network can memorize about 0.15 patterns per neuron with correlation learning, the past research results in our laboratory shows that it can memorize more than 0.7 patterns per neuron with incremental learning. So incremental learning is appropriate for associative memory, and to find the cause of that is the purpose of this research.
In this research, we hypothesized that the abirity comes from two causes. The first was to use the weights which had formed when the neural network had learned former pattern in learning unknown pattern. The second was adequacy of the weights distribution which came from the great number of the value of the weights.
From the results of this research, it was verified that the neural network uses the weights which had formed when it had learned former pattern in learning unknown pattern. And, it was considered that the value of the weights doesn't have to be many if its distribution was appropriate.