Barchi Library (140 John Morgan Building)
Department of Physics and Astronomy
Mapping neural signals to text and Fluctuation analysis of the inner ear’s active dynamics
An important topic in current sensory neuroscience research is studying the production and audition of speech. In this regard, I will elucidate my contribution in a brain-computer interface project, wherein we decode text directly from neural signals underlying speech production. While the eventual goal of this project is to create an assistive device for patients suffering from debilitating neuromuscular diseases, our work lies in its open-loop offline counterpart. We implement a framework that initially isolates frequency bands in the neural recordings encapsulating differential information of various phonemic classes. These then form the input feature set for an LSTM which at each time point outputs phoneme probabilities that are eventually fed into a particle filtering algorithm previously trained on the Brown corpus to generate the resultant textual word.
Additionally, I will allude to my work on active dynamics of the inner ear. The inner ear lies at one of the early stages in the auditory detection of sound and is an active non-linear system capable of parsing pressure waves ranging over several orders of magnitude in frequency and amplitude. Its mechanically sensitive hair cells, named for the bundles of stereocilia protruding from their apical surfaces, can even detect sounds that result in Ångstrom-scale displacements. While a significant body of work indicates that an internal active mechanical process serves to amplify the incoming signal, biophysical mechanisms behind this acute sensitivity are not fully understood. This active motility also governs spontaneous limit cycle oscillations by these stereociliary bundles demonstrated in vitro by a number of species. Present-day theoretical models that study this bundle motility are complex and expansive compared to experimental observations which are restricted to only a few of the variables, rendering the models highly susceptible to over-fitting. In this context, through my work I suggest a framework to test and enhance the fidelity of these models to data via the formalism of fluctuation analysis of stochastic bundle motion.
A pizza lunch will be served.