Varona-Moya, S. & Cobos, P. L.
Department of Basic Psychology, Faculty of Psychology, University of Málaga (Spain)
In the well known paper ``Learning Distributed Representations of Concepts´´, Hinton accomplished a pioneering study of multi-layer perceptrons' capacity to make analogical inferences in the so-called family trees task. He exposed a five-layer feed-forward network to items that were not part of its training set (i.e., four relationships between members of two genealogically identical families), to see whether it could nevertheless generalize and produce the correct output. The reported results of two simulations were very good and it was concluded that this type of network could detect the common structure shared by both families and therefore make analogical inferences from one family tree to the other. However, some critical aspects regarding the way in which those results were obtained cloud their interpretation and make advisable to review that ground-breaking work. Specifically, we claim that it lacked a proper control condition to ensure that the test items were not solved through strategies other than analogy-based generalizations and also that unsuitable test items that did not actually induce such inferential processes were very likely used. Thus, we conducted a series of simulations in which these two aspects were tackled through experimental manipulation. Our results, based on 500 simulations, confirmed that the good performance of Hinton?s simulations was probably an artifact, but nonetheless proved that his perceptron is actually able to generalize knowledge between structurally identical domains, providing a more accurate assessment of such capacity.