On Computational Evidence for Different Types of Spatial Relations Encoding: Reply to Cook Et Al. (1995)

Computational models in psychology play an increasingly important role in characterizing theoretical distinctions, understanding empirical results, and formulating new predictions. However, the proper use of models is subject to debate and interpretation, as Cook, Früh, and Landis (1995) have demonstrated in a critique of neural network simulations reported by Kosslyn, Chabris, Marsolek, and Koenig (1992). These simulation results supported a distinction between two types of spatial relations encoding. Cook et al. argue that Kosslyn et al.'s models did not process "spatial" representations and that input-output correlations rather than properties of spatial relations encoding processes explain the performance of the models. This article provides conceptual and analytic rebuttals of those criticisms.