The cerebral hemispheres and neural network simulations: Design considerations.

Abstract

The conclusions concerning hemispheric specializations based on neural network simulations, which were previously reported by Kosslyn, Chabris, Marsolek, and Koenig (1992), are shown not to be valid. Differences in network performance on tasks said to be "categorical" and "coordinate spatial" in nature were due to imbalances in the input stimuli and cannot, in principle, be related to differences in performance on such tasks in human subjects. The use of truth tables and correlation coefficients in the design of neural networks is discussed.

Encoding Categorical and Coordinate Spatial Relations Without Input‐Output Correlations: New Simulation Models

“…And incidentally, the fact that in our tasks the categorical mapping contains higher input-output correlations but is nevertheless more difficult shows that input-output correlations alone cannot explain network performance. Cook et al (1995) have reminded us that it is always important, especially when using neural-network simulations, to ensure that one's experimental design can rule out altemative hypotheses. Now that we have ruled out a trivial explanation for the effects Kosslyn et al (1992) found, we have produced more substantial evidence supporting the computational distinction between processes that encode categorical and coordinate spatial relations representations.…”

“…A graph of the cp/cp,, input-output correlations for each input unit and the output units that represent each task is shown in Figure 2. Cook et al (1995) argue that because some of these correlations are 1 .O and many of the others are high, much of the networks' performance can be attributed to these trivial lirstorder input-output correlations, and the various results discussed by Kosslyn et al (1992) cannot with confidence be interpreted as relevant to processes that compute spatial relations representations. In other words, because of the specific training set used, the networks may have learned by "cheating" rather than by developing a computational structure appropriate to the task that humans perform, and so "conclusions of interest to human psychology cannot be drawn" from examining these networks (Cook et al, 1995, p. 420;Cook, 1995, makes a similar point in reference to Jacobs & Kosslyn, 1994).…”

“…Thus we now present the results of new simulations in an attempt to examine the issue. Cook et al (1995) raise the point that because the bar portion of the input patterns used in the neural-network simulations in Kosslyn et al (1992) could fall in only five locations near the center of the input array, the training set contained "definitive information" that trivialized the computation the network performed. For example, any time one of the upper nine units was active, the dot was above the bar, and therefore the correct response in the categorical task was "above."…”

Neural Network Models as Evidence for Different Types of Visual Representations

“…Cook (1995) has now focused on the issue of input-output correlations with respect to the simulations reported by Jacobs and Kosslyn (1994). In response to Cook et al (1995), Kosslyn et al (1995) addressed this last criticism in detail (see pp. 427-429), so we will not repeat that discussion here.…”

“…Cook and his colleagues have criticized these simulations on several grounds. In particular, discussing the findings of Kosslyn et al (1992), Cook, Friih, and Landis (1995) argued that neural network models do not process "spatial" information, and that even if they do, the Kosslyn et al models were flawed because their training patterns contained so-called "definitive information," which made it possible for the networks to encode spatial relations without developing generalizable representations. Cook (1995) has now focused on the issue of input-output correlations with respect to the simulations reported by Jacobs and Kosslyn (1994).…”