| How do people acquire, represent, and process the information that allows them to perform semantic tasks? To answer this question, many theorists adopt two tenets: first, that semantic memory is mediated by a process of categorizaiton; and second, that such processes are grounded in a system of implicit theories which explain object-property correspondences. In my dissertation, I will argue that this may not be the optimal approach. Instead, I suggest that the semantic system exploits more graded representations, which emerge as domain-general learning mechanisms assimilate the structure of the environment. From this view, learned similarities among object representations constrain how knowledge generalizes, how easily it is acquired, how quickly it can be accessed, and how it degrades in dementia. I will illustrate these principles with reference to two computational models, which represent complementary implementations of the theory. The first is a feed-forward model which learns the mappings between objects and their properties in different contexts. Computer simulations with the model illustrate principles that explain a range of disparate empirical phenomena, including data from infant preference studies, lexical acquisition, basic-level primacy, expertise, semantic dementia, and theory theory. The second is a recurrent model which acquires the mappings among visual representations of objects, their names, and their verbal descriptions. When the model's inputs and outputs are constructed to match representative data from attribute-listing and drawing data, it acquires internal representations whose similarity structure captures the semantic relations among objects. I compare the model's performance under simulated lesions to data from neuropsychological patients with generalized semantic deficits. The structure revealed in impaired patient performance is interpreted as arising from representational similarity structure in the semantic system, which in turn emerges from the domain-general learning mechanisms embodied in the model. |
