Varona-Moya, S. & Cobos, P. L.
Department of Basic Psychology, University of Málaga, Málaga, Spain
Multilayer perceptron networks’ ability to perform sensible inferences by analogy was pointed out by Hinton from the results of knowledge generalization tests between two isomorphic family trees. Due to its importance, a comprehensive review of this work was tackled to improve some methodological aspects in order to find statistically grounded answers to the questions posed by the author. Using the same network architecture and learning procedure, 500 simulations were trained in the task proposed by Hinton. In this review (1) the degree of isomorphism grasped by a simulation was computed through an ad hoc algorithm applied to principal component analysis scores of hidden units’ activation vectors and (2) the relationship between the network’s ability to perform inferences by analogy and its grasp of isomorphism was examined on a mixed factorial design basis, using corrected generalization tests. The main conclusions are these: (1) isomorphism grasp is not as consistent a property as Hinton suggested, since many simulations failed to build identical representations for both family trees, and (2) statistically significant interaction effects between isomorphism grasp and learning the second family tree proved that the more isomorphism a simulation grasped, the better it generalized from one tree to another.