The design and implementation of novel computational and machine learning approaches for modelling brain dynamics: towards more interpretable and real-time brain analysis
File(s)
Author(s)
Dinov, Martin
Type
Thesis
Abstract
This thesis presents a combination of novel methods intended for improving Brain Computer Interface (BCI) use, such as a Dynamic Time Warping-based (DTW) spectrum, as well as new applications of existing methods, such as fuzzy clustering and neural networks, including reinforcement learning-driven Deep Q Networks (DQNs). We develop these mutually-compatible methods that aim to make brain analysis, especially in the context of BCIs, more interpretable and efficient.
In Chapter 2, I developed a new approach based on using DTW towards computing frequency-domain spectra in a more interpretable way than using standard Fourier or Wavelet spectrums. Though it is applicable to any time series data, I applied the DTW-spectrum to EEG and show that it explains more variability in brain dynamics compared to other standard measures - most notably it seems to better predict certain benchmark measures of brain dynamics than the corresponding Fourier transforms. Chapter 3's main topic is using the fuzzy c-clustering and softmax neural network-based fully probabilistic classification and analysis framework that I developed for EEG microstates, which shows significant issues with standard deterministic analyses that have been used heretofore. Using a large publically available data set, I showed that imagined motor movements are less predictable than real motor movements. Further, I suggest that treating microstates as states representing a discrete dynamic of the brain is losing valuable information regarding the underlying dynamics. Chapter 4 is focused around Reinforcement Learning (RL)-driven Behavioral- and Neuro-Feedback (NFB) using phasic auditory alerts. This proof of concept work shows simulation and experimental results that suggest that the DQNs can learn meaningful behaviors with a portable consumer EEG within a small number of trials.
The work necessitates a somewhat meandering journey through the relevant neuroscientific, mathematical and computational literature, which is covered by the background and foundation laid out in Chapter 1.
In Chapter 2, I developed a new approach based on using DTW towards computing frequency-domain spectra in a more interpretable way than using standard Fourier or Wavelet spectrums. Though it is applicable to any time series data, I applied the DTW-spectrum to EEG and show that it explains more variability in brain dynamics compared to other standard measures - most notably it seems to better predict certain benchmark measures of brain dynamics than the corresponding Fourier transforms. Chapter 3's main topic is using the fuzzy c-clustering and softmax neural network-based fully probabilistic classification and analysis framework that I developed for EEG microstates, which shows significant issues with standard deterministic analyses that have been used heretofore. Using a large publically available data set, I showed that imagined motor movements are less predictable than real motor movements. Further, I suggest that treating microstates as states representing a discrete dynamic of the brain is losing valuable information regarding the underlying dynamics. Chapter 4 is focused around Reinforcement Learning (RL)-driven Behavioral- and Neuro-Feedback (NFB) using phasic auditory alerts. This proof of concept work shows simulation and experimental results that suggest that the DQNs can learn meaningful behaviors with a portable consumer EEG within a small number of trials.
The work necessitates a somewhat meandering journey through the relevant neuroscientific, mathematical and computational literature, which is covered by the background and foundation laid out in Chapter 1.
Version
Open Access
Date Issued
2017-11
Date Awarded
2018-04
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Leech, Robert
Sharp, David
Sponsor
Great Britain. Ministry of Defence
Defence Science and Technology Laboratory (Great Britain)
Grant Number
DSTLX1000083275
Publisher Department
Department of Medicine
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)