Achievement

The relevance to human cognition of abstract descriptions of language processing with symbol-manipulating rules has been questioned based on a type of computational neural network model. But trainee Christo Kirov has shown that a model of this type is in fact understandable as an implementation of such rules. The symbolic account of dependencies between verbs and their subjects in certain languages like Dutch posit two types of basic symbolic memory systems. Roughly, sentences analogous to 'boys dogs doctors inoculate bite survive' require retrieving the *last* noun to serve as subject when the first verb is encountered; 'boys dogs doctors survive bite inoculate' requires the *first* noun instead. Kirov trained a neural network to store serially-presented symbols and then retrieve them either last-first or first-first. His analysis revealed important evidence that the network had learned to implement with activation values the two separate memory systems of the symbolic account.