Human cognitive neuroscience seminar

Speakers: Guillermo Puebla (University of Edinburgh)

Topic: The processing limits of Parallel Distributed Processing Models: Relational generalization and statistics independence

Abstract: The ability of traditional Parallel Distributed Processing (PDP) models to capture relational knowledge is a matter of long-standing controversy. Recently, Rogers & McClelland (2008, 2014) have contended that the Story Gestalt model of text comprehension is capable of binding arguments to relational roles. The present study evaluates the Story Gestalt model in two core capacities of relational thought: relational generalization and statistics- independent inference. Regarding relational generalization, we found that, when trained in a highly combinatorial corpus, the model is able to correctly process new stories composed of known elements, but is unable to correctly process stories where a new concept plays a trained relational role. Concerning statistics-independent inference, we found that the model is unable to correctly process stories that violate the statistical regularities of the training dataset. Importantly, we replicated these results in versions of the model that use both localist and distributed representations of concepts. We argue that relational generalization and statistics-independent inference capabilities of the model are limited because it cannot perform dynamic binding of independent roles and fillers. Ultimately, these results cast doubts on the suitability of the PDP framework for explaining phenomena based on relational knowledge.