Virginia Savova, Ph.D.

Current research: vision

I study aspects of visual learning and generalization involving complex structured objects. My work focuses on visual categories exhibiting high degrees of internal variation within abstract structural constraints. Unlike other well-studied object types (e.g. faces), complex structured objects (e.g. houses) are not well-captured by template-based recognition models, which assume a rigid set of elementary subparts. We use Bayesian inference on grammar-based representations to explain the rapid and subconscious knowledge acquisition concerning these objects in the visual domain.

Coming soon

“Grammar-based representations in visual scene parsing.” Submitted.

For published work see my Publications page.

Bio sketch

I was born in Sofia, Bulgaria, in 1976. I arrived to the United States in 1995 to begin my undergraduate studies at Harvard in Linguistics. I went on to earn a doctorate in Cognitive Science from the Johns Hopkins University. I have worked and published in linguistics (minimalism, optimality theory, formal language theory), psycholinguistics, natural language processing, and philosophy of AI. I am currently developing and applying Bayesian models to the acquisition of rich structural representations in vision and language.

Postdoctoral Associate

Computational Cognitive Science Lab

Department of Brain and Cognitive Sciences

Massachusetts Institute of Technology

Current research: language

I study the role of syntactic optionality in language, particularly as it relates to word-order variation. Previous research has shown that non-canonical word order increases processing time, due to added structural complexity. In that case, why is non-canonical word order used at all? We hypothesize that non-canonical word order exists because it realizes discourse linearization preferences. For example, it is advantageous from a discourse point of view to introduce established discourse entities before novel ones. If the canonical word order does not allow for this possibility, the non-canonical word order is expected, and processed easily by the recipient.

In preparation

“Structured probabilistic models of the interaction in discourse and syntax on-line reading.”

For published work see my Publications page.