Comparing two methods of visual abstraction to human data

The ability to “visually abstract” a given pattern with a neural network and abstract the same pattern by using a regression/correlation analysis was investigated. Both methods were compared with human subjects performing the same task. To visually abstract a particular shape, both quantitative methods broke the shape down into its linear, quadratic, and cubic components. Using an IBM-compatible personal computer, 10 test patterns were analyzed with a neural network (designed using Brainmaker Professional and trained with known linear, quadratic, and cubic shapes) and a regression/correlation model (designed using Lotus 1-2-3). The 10 test patterns were also analyzed by 22 human subjects. The neural network data were found to be highly correlated with the human data [r(8) = .90,p < .01]. The regression/correlation model’s data were also found to be significantly correlated with the human data [r(8) = .77,p < .01]. These findings demonstrate the successful modeling of Rumelhart’s (1991) regression/correlation approach to visual abstraction.