Recall time in sparsely encoded Hopfield-like associative memory

Recall time in sparsely encoded Hopfield-like associative memory under parallel dynamics is investigated on the basis of computer simulation. It is shown that the recall time is power dependent on the network size. Furthermore, the dependence of power index on sparseness, information loading and distance between an initial network state and the recalled prototype is investigated. The power index decreases when information loading and the initial distance from the recalled prototype decreases. It essentially decreases when sparseness increases. From these results, traditional and neural network approaches for pattern matching tasks are compared. It is shown that for extremely complex tasks, when an entropy of a signal space H is in the order of more than hundreds and the number of stored patterns L≫H, the neural network approach outperforms the traditional one in processing rate even if it is running on a normal serial computer