*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Brianna Karpowicz
BME PhD Proposal Presentation
Date: 2022-10-06
Time: 3:00 pm (ET)
Location / Meeting Link: https://emory.zoom.us/j/97608060343
Committee Members:
Chethan Pandarinath, PhD (Advisor); Hannah Choi, PhD; Lee Miller, PhD; Lena Ting, PhD; Anqi Wu, PhD
Title: Stabilizing brain-computer interfaces using deep-learning dynamical systems models
Abstract:
Intracortical brain-computer interfaces (BCIs) may restore abilities to paralyzed individuals by monitoring their brain activity and mapping it to an external variable using an algorithm known as a decoder. However, decoders are subject to instabilities at the neural interface, which can cause changes in the precise neurons being recorded. This changes the relationship between recorded neural activity and behavior, leading to massive degradations in decoding accuracy. Standard solutions require interrupting device use to collect new data and recalibrate the decoder. To maintain high-accuracy decoding without interruptions to device use, it is necessary to develop a recalibration procedure that does not require the collection of new behavioral data. In the long term (weeks to months), one promising approach leverages the latent manifold structure that captures patterns of co-activation across neurons, which maintains a consistent relationship to behavior over time. This allows the same decoder to be used as long as we can map new neural data onto the manifold. Using dynamics – time-varying patterns of neural population activity - to uncover the manifold has led to better decoding accuracy. In Aim 1, we test whether using neural dynamics in mapping neural activity onto the manifold can improve the stability of BCI decoders. A viable solution also requires that we are able to predict behavior accurately in the short term (hours to days) so that we can collect sufficient data and learn to update the mapping to the manifold. In Aim 2, we will combine local field potentials with the short-term stability conferred by self-supervised learning to build self-supervised, multiscale dynamics models and assess their impact on decoding stability. In Aim 3, we plan to develop a benchmark for standardized comparison across stabilization techniques.
