Emergence in Chinese Character Learning

The formation of concepts is a key point in development of human intelligence. There are two kinds of traditional ideas about this problem: hypothesis elimination and associative learning. The Bayesian framework suggested by Tenenbaum absorbs the virtue and avoids the limitations of the two traditional ideas and its results are demonstrated by behavioral experiments. The model simulates statistics behavior of human concept learning but does not refer neural mechanism of concept learning. We try to simulate neural behavior during the process of formation of concept by Hopfield-type neural network and suggest that concepts could correspond to stable attractors of the system. In results we can see that the network could learn Chinese characters quickly and its weights will change gradually to approach fixed values finally as emergence. We find that the Hebbian learning rule adjusting weights of our network could describe the different action of the two traditional ideas-filtering and abstraction and guess it may be a kind of microcosmic mechanism of formation of concepts in certain conditions.