Miller´s magical numbers are relevant to neural networks

G.A. Miller (1956) analyzed psychological evidence that quantifies human capability to discriminate sensory signal inputs and to perform short-term recall given symbolic inputs. He found that both types of experiment were characterized by a small memory capacity on the order of seven. The authors reexamine Miller´s paper and data from an artificial neural network perspective to gain insight into discrimination capacity as a function of input signal space capacity. It is concluded that the key to robust discrimination capability in humans does not rely on either high numerical precision or complex mappings of input states to output states, two common characteristics of many artificial neural networks. The capability does rely on large numbers of neurons, again at variance with common practice