Simulating Event-Related Potential Reading Data in a Neurally Plausible Parallel Distributed Processing Model - eScholarship
Sarah Laszlo, David C. Plaut · Proceedings of the Annual Meeting of the Cognitive Science Society · 2011
Simulating Event-Related Potential Reading Data in a Neurally Plausible Parallel Distributed Processing Model Sarah Laszlo ([email protected]) Department of Psychology and Center for the Neural Basis of Cognition, 5000 Forbes Ave Pittsburgh, PA 15213 USA David C. Plaut ([email protected]) Department of Psychology and Center for the Neural Basis of Cognition, 5000 Forbes Ave Pittsburgh, PA 15213 USA Abstract Parallel Distributed Processing (PDP) models have always been considered a particularly likely framework for achieving neural-like simulations of cognitive function. To date, however, minimal contact has been made between PDP models and physiological data from the brain performing cognitive tasks. We present an implemented PDP model of Event-Related Potential (ERP) data on visual word recognition. Simulations demonstrate that a novel architecture with improved neural plausibility is critical for successfully reproducing key findings in the ERP data. Keywords: Parallel Distributed Processing (PDP), Event- Related Potentials (ERPs), visual word recognition Introduction From their initial development, PDP models have been considered an especially promising framework for building simulations which perform cognitive tasks with a mechanism similar to that employed in the brain (e.g., McClelland, Rumelhart, & Hinton, 1986). This optimism derives in large part from the fact that the basic processing units in PDP models are neuron-like, in that the models typically employ many interconnected units, each performing relatively simple computations, and represent information in a distributed fashion (c.f., Bowers, 2009; Plaut & McClelland, 2010). The sense that PDP models should lend themselves well to simulating data from cognitive neuroscience—that is, brain data relating to cognitive function—is not only historical. Indeed, especially in the domain of single word reading, it is currently common for descriptions of prominent models to suggest that improvements over existing models could and should be made by increased contact with data from cognitive neuroscience (e.g., Harm & Seidenberg, 2004; Perry, Ziegler, & Zorzi, 2007). Correspondingly, as theories of how reading works based on neuroimaging data have become increasingly well- specified, a consensus is emerging—especially in the Event-Related Potential (ERP) literature—that interpretation of brain data could benefit from the guidance of formal computational models (e.g., Banquet & Grossberg, 1987; Barber & Kutas, 2007; Van Berkum, 2008). For example, one currently viable theory of the functional significance of the N400 ERP component (a centro- posterior component peaking around 400 ms post stimulus onset, and thought to reflect lexical-semantic access: see Kutas & Federmeier, in press, for review) suggests that N400 activity represents the continuous activation of semantic features associated with an orthographic input at either a whole or partial item level (e.g., the activation of the semantic features associated with both FORK and PORK in response to presentation of the word FORK; Laszlo & Federmeier, 2011). Under this so called obligatory semantics view, contact with semantics is made automatically by every orthographic input, and interaction between levels of representation is continuous (explaining, for example, sentence context effects on illegal nonwords; Laszlo & Federmeier, 2009). Two features of this theory are particularly relevant for implementation in a computational model. First, the proposal that orthographic sub-parts of items can activate the semantic features of orthographically similar items extends to nonwords, such that pseudowords (e.g., GORK) and even consonant strings (e.g., XFQ) are allowed to contact semantics —explaining robust N400 effects observed for these items (e.g., Laszlo & Federmeier, 2009). This feature of the obligatory semantics view implicates a word recognition system that is not strongly lexicalized, and as such would seem to be more appropriate for simulation in a distributed PDP framework than in competing frameworks with explicit lexical representations (cf., Perry, Ziegler, & Zorzi, 2007). Though it would be possible for lexicalized models to account for these data by simply allowing very un-wordlike nonwords to activate their neighbors weakly, such a system is no longer strongly lexicalized in that its internal response to each input involves the activation of a number of units, with that activation graded by similarity to the input-- a system that is essentially distributed. Second, the continuous, interactive nature of the obligatory semantics view stronly contrasts with staged models of word recognition (e.g., Borowsky & Besner, Thus, the obligatory semantics view posits a mechanistic account of visual word recognition resonant with the PDP approach, but the question remains: would an implemented PDP model exhibit the patterns of effects in the ERP data suggestive of a non-lexicalized, continuous system (e.g., N400 effects for illegal consonant strings)?