Slide Slam K12
Modeling the mental lexicon using a spiking neuron network for simulating speech tasks in case of different types of aphasias
Bernd J. Kröger1, Trevor Bekolay2, Terrence C. Stewart3; 1RWTH Aachen University and UKAachen, Aachen, Germany, 2Applied Brain Research, Waterloo, ON, Canada, 3National Research Council of Canada, University of Waterloo Collaboration Centre, Waterloo, ON, Canada
Introduction: Large-scale neural models allow to set up of a concrete architecture for all functional modules of speech processing and give a concrete temporal sequencing of neural activations occurring during speech production or speech perception tasks. These models allow us to develop a better understanding of the basic functioning of neural processes occurring in speech processing. Method: Our neurocomputational model is based on the NENGO.ai-approach. The model in its current version is capable of simulating word production tasks (e.g., picture naming), word comprehension tasks (e.g., finding a superordinate concept or generic term) as well as word repetition tasks. Different versions of the model have been implemented for modeling a normal speaker as well as six types of speakers suffering from different types of aphasias, i.e. Broca’s, Wernicke’s, transcortical motor, transcortical sensory, conduction and mixed or global aphasia. In case of each of the model variants about 20 different speakers were modelled, suffering from different degrees of neural dysfunctions representing the appropriate type of aphasia. All three types of tasks (production, perception, repetition) were applied to each of the model speakers and each model speaker performed 3 tasks each, leading to about 350 simulation runs. Results: By analyzing the task performance of all runs we were able to generate typical symptoms of aphasic speech. Performance rates decreased differently with increasing degree of neural dysfunction for the production, perception and repetition task in case of the different types of aphasia as predicted by natural data. Conclusions: The model gives new insights, how speech processing can be understood from a neuro-functional perspective because the neural model used here is neurobiologically plausible: (i) The model is based on a functional network of about 200000 spiking neurons (LIF neurons); (ii) the model uses neuron buffers containing neurobiologically plausible neural activation patterns for representing phonological, lemma or semantic forms; (iii) the model uses neurobiologically realistic associative memories in order to simulate transformations of neural activation patterns from concept via lemma to phonological form levels and vice versa; (iv) the model comprises an action control component (cortico-cortical control loop including a model of the basal ganglia and of the thalamus) for modeling a correct temporal sequencing of all peripheral and central actions needed to simulate a speech task. Thus, the neural architecture of the model and the temporal succession of neural activation patterns occurring within different parts of the model elucidate basic functional principles of the interaction of the mental lexicon (central knowledge repository) with different neuron buffers representing neural activation patterns at different levels of the production and perception pathway.