PURPOSE: To increase the resistance to the noises or to store the similar elements of a time series while identifying them with each other by using the generalizing ability of a neural network to absorb the fluctuation of the values of a retrieving key and each element of the time series.

CONSTITUTION: In a retrieving mode, a retrieving key set to a time series is given to a context layer 54 with a start mark given to an input layer 51 as the initial state for output of a 1st element of the time series. At the next time point, the value set to an intermediate layer 52 precedently by one time point is shifted to the layer 54. Then the value set to an output layer 53 precedently by one time point is given to the layer 51. Thus the output of a recurrent network is obtained. This operation is repeated until the output of an end mark so that the elements of the time series are successively outputted. In such a way, the correspondence is secured between the retrieving key and the time series as well as between the elements of the time series with the associative storage attained via a neural network. Thus it is possible to secure the associative storage that has the high resistance to the noises and can identify the similar elements of the time series with each other.