Linear and non-linear methods of source separation in neurophysiological time series
File(s)
Author(s)
Clarke, Alexander Kenneth
Type
Thesis
Abstract
Blind source separation of neurophysiological time series into its constituent neural activity is a key component of modern concepts of neural interfacing and an important tool for motor neuroscientists. With concurrent rapid advances in high-density electrode arrays, there is now a real possibility of devices driven directly by the nervous system leaving the lab and becoming a core part of machine interaction in everyday life. Such an advance would represent a leap forward in a diverse range of technologies, including intuitive prosthetics, neurological diagnostics, and consumer device. However, contemporary linear algorithms can return poor results when subjected to some of the signal contaminants that will be faced in the wild, such as high or varying noise levels and the non-stationary effects of dynamic contraction. New approaches are needed if source separation routines are to be robust enough for practical use.
One clear avenue of enquiry can be found in the recent explosion in techniques for training non-linear functions such as deep neural networks. Deep neural networks are highly flexible and have proven efficacy on a variety of data types, including many time series applications. However, their adoption would significantly increase the complexity of the blind source separation pipeline, so their presence would need to be well justified. In this thesis I explore a range of deep learning methodologies, and in particular how they can be effectively blended with contemporary linear algorithms. I demonstrate that such hybrid approaches can bring a range of advantages, such as improved management of noise, the automatic identification of incorrect timestamps and the ability to compensate for non-stationary effects in the signal.
One clear avenue of enquiry can be found in the recent explosion in techniques for training non-linear functions such as deep neural networks. Deep neural networks are highly flexible and have proven efficacy on a variety of data types, including many time series applications. However, their adoption would significantly increase the complexity of the blind source separation pipeline, so their presence would need to be well justified. In this thesis I explore a range of deep learning methodologies, and in particular how they can be effectively blended with contemporary linear algorithms. I demonstrate that such hybrid approaches can bring a range of advantages, such as improved management of noise, the automatic identification of incorrect timestamps and the ability to compensate for non-stationary effects in the signal.
Version
Open Access
Date Issued
2024-06-23
Date Awarded
01/02/2025
License URL
Advisor
Farina, Dario
Bentley, Paul
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/L016737/1
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)