Dissertation Defense - Leslie Famularo / Grounding speech perception modeling in auditory neuroscience through differentiability
Dissertation Defense - Leslie Famularo / Grounding speech perception modeling in auditory neuroscience through differentiability
March 12, Leslie (Li) Famularo defends her dissertation, "Grounding speech perception modeling in auditory neuroscience through differentiability," abstracted below, in Room 3137 of the Iribe Center.
Newborns are sensitive to the difference between the speech of some languages but not others, a phenomenon referred to as early language discrimination. While this is commonly attributed to their sensitivity to the temporal rhythm in speech, it has never been systematically tested. In this thesis, I explored the behavioral phenomenon of
language discrimination through a series of simulations using machine learning models. In addition to typical models directly drawn from machine learning, I also introduced a model that is grounded in auditory neuroscience through differentiable programming.
Results from the traditional machine learning models suggest that rhythm was not necessary for any model to perform language discrimination in a humanlike manner, which implied that other mechanisms relying on global statistics alone could be possible for language discrimination and potentially used by humans during behavioral tests. Additionally, with the differentiable model with auditory neuroscience constraints, while the model uses rhythm to
perform language discrimination, the range of rhythm was much faster than what is associated with syllable rhythm. These results have implications about newborn language perception and language acquisition that follows, and may be used to drive the design of future infant studies. Additionally, the application of differentiable programming to introduce intuitions and constraints from neuroscience and cognition offers a new path of manipulating deep neural networks in the study of neural and cognitive modeling.