A Parallel Distributed Processing Network of Human Reasoning

Humans are found to display a bias when they are doing deductive reasoning in the Wason selection task (WST). Specifically, when asked to identify evidence that disproves a rule (e.g., that a card with an even number on one side must contain a vowel on the opposite side), humans perform better when the task uses real-life examples than abstract examples. The mechanism of this reasoning bias remained unclear. One study used a parallel distributed processing (PDP) model to model the human-like bias but failed. The possible reason for the failure was that semantic similarities between words/concepts used in the task could induce the bias, and the semantic similarity among learned concepts was not simulated in the PDP model. Regardless, PDP offers an exciting avenue to examine decision making due to its theoretical and biological plausibility. And so, the current study used a PDP approach to model the deductive reasoning behaviors in the Wason selection task. Different from the previous PDP model study, we trained the PDP model with a WST under two semantic distance conditions: close and far. We found that the PDP model learned to choose “p and not q” under the close semantic distance condition and “p and q” under the long semantic distance condition. However, unlike human participants, the PDP model did not find it more difficult to choose “p and not q” over “p and q” in either condition. Future studies might explore how the model’s cause-effect experiences without proper feedback prior to exposure on the Wason selection task may produce biased outcomes.