111. deep learning methods for analyzing neural data
Name: FNU Pailla-Tejaswy
Grad Year: 2018
Motor brain-machine interfaces(BMIs) map neural activity to volitional movement kinematics and can be used as assistive devices for people affected with stroke, tetraplegia and other neurodegenerative diseases. Designing robust BMIs requires us to capture informative patterns from non-stationary and often noisy neural signals. Traditionally, spectral powers from specific frequency bands coupled with mostly linear decoding models have been used for BMI design. We propose using deep neural networks, with architectural design decisions guided by domain knowledge, to capture spatio-temporal patterns in neural data. This would simplify BCI design by encapsulating feature engineering and model selection in one framework. In the current work, we use deep neural networks to decode finger movements from electrocorticographic data. We investigate the time-frequency patterns learnt by the networks and discuss how they relate to the current knowledge of neuroscience. With an increasing interest in high density recording systems, the quantity of neural data recorded is set to increase exponentially. Our methods will aid neuroscientists in mining large scale neural data and alleviate the task of hand-designing features and decoding models for neural engineers.
Industry Application Area(s)
Life Sciences/Medical Devices & Instruments | Software, Analytics