Introduction to special issue on computational modelling in cognitive neuropsychology

Abstract

Recently, Cognitive Neuropsychology altered the statement of the journal’s aims that appears in every issue, the new material expressing an interest in publishing “computational modelling research that is informed by consideration of neuropsychological phenomena”. The journal’s interest in computational models, however, is anything but recent. When Cognitive Neuropsychology was in its first decade, it devoted an entire issue to a single article reporting a model of impaired reading by Plaut and Shallice (1993). This now classic treatment of dyslexic errors from a connectionist perspective—it is currently the third most cited article in the journal—and later examples of what Coltheart (2006) calls “computational cognitive neuropsychology” have influenced the interpretation of neuropsychological data for some time now. This special issue thus comes at a time when a variety of computational models and modelling frameworks have had their say on neuropsychological data. What is a computational model? It is a model expressed as a computational implementation, the implementation being necessary to understand the model’s implications for data and for theory. What, then, is a model? The term “model” has been used in the context of accounts of impaired cognition since the inception of the field. For example, the Wernicke–Lichtheim model of aphasia (e.g., Lichtheim, 1885) was a diagram containing nodes representing mental content (e.g., auditory lexical images) and directional arrows indicating the flow of processing among the nodes during various tasks (e.g., speech comprehension). Lesions could be associated with nodes or arrows, and, because each of these components was identified with both a cognitive function and a brain region, the model generated accounts of aphasic symptoms—how they clustered and how they were associated with brain areas. For example, the poor comprehension and repetition associated with Wernicke’s aphasia was attributed to a lesion to the auditory-image node. The diagram revealed that this node is a necessary part of the path for the tasks of both comprehension and repetition. Thus, the model, in the form of Lichtheim’s diagram, made concrete the links between the theory and potential patterns of data. Modern computational models are, of course, much more than diagrams. But they have much the same function. The model’s computer program specifies a cognitive architecture, including processing levels or subsystems, and a variety of operations and parameters. This program is analogous to the diagram. Running the program simulates the cognitive processes involved in the performance of some task, just as following the arrows does in a diagram. Of course, the program is necessary because the complexity of most computational models precludes working things out by hand. Various aspects of the program can be