A computer model of infant perceptual development

Abstract

A theory is presented of the development of pattern recognition and looking behaviour in infancy. It is proposed that scanning habits are acquired and patterns recognized with the reproduction of fixations and eye movements in the order in which they originally occurred. Recognition is achieved by correctly predicting the current input for each fixation. Evidence supporting this proposal is discussed, and the limitations of other theories are examined. A case is made for the storage of two kinds of visual information, originating from central and peripheral vision respectively. Infants indicate recognition of familiar patterns by looking less at them than patterns which are new. This can be explained by the discrepancy principle which proposes a curvilinear relation between the amount of looking and degree of discrepancy between a pattern and its representation in memory. This principle is incorporated in the theory to account for the control of the length of sequences of fixations. A computer model of the theory is described. This contains a simulation of the cortical processing of visual input, a number of oculomotor reflexes, learning mechanisms, and the means of controlling the length of a fixation sequence by assessing its discrepancy with the contents of memory. The model was run on a computer and learned to recognize patterns by scanning them and reproducing the original sequences of fixations. The ability of the model to mimic infant looking behaviour is shown in three simulations of different infant experiments. Recognition was demonstrated by a decline in looking at familiar relative to new patterns, and this ability was retained after a delay. Such behaviour took time to develop, and the model required a certain level of visual experience before it appeared. Individual differences in the performance of the model resembling tempo differences in infants were also produced.