Abstract

Cook (1995) criticized Kosslyn, Chabris, Marsolek & Koenig's (1992) network simulation models of spatial relations encoding in part because the absolute position of a stimulus in the input array was correlated with its spatial relation to a landmark; thus, on at least some trials, the networks did not need to compute spatial relations. The network models reported here include larger input arrays, which allow stimuli to appear in a large range of locations with an equal probability of being above or below a “bar,” thus eliminating the confound present in earlier models. The results confirm the original hypothesis that as the size of the network's receptive fields increases, performance on a coordinate spatial relations task (which requires computing precise, metric distance) will be relatively better than on a categorical spatial relations task (which requires computing above/below relative to a landmark).