My research focuses on Brain-computer interfaces (BCIs) and how they can be used to provide a unique perspective into motor learning processes and the interactions between the brain, body, and environment. I am also very interesting in broader neuroscience questions such as non-linear dynamics and measures of functional connectivty in the brain, for example, phase synchronization and complex network methods. My research used electroencephalographic signals (EEG), functional Near infrared spectroscopy (fNIRS), and functional Magnetic resonance imaging (fMRI).
Brain computer interfaces (BCIs)
BCIs allow control of a computer directly from signals recorded from the central nervous system. As such they allow individuals to control computers via brain activity alone without the need for use of the efferent nervous system or muscles. This makes BCIs a very useful tool, allowing individuals who suffer from severe motor impairments to learn to control some aspects of their environment and communicate with their friends, family and care givers
BCIs have traditionally been developed for use by individuals with severe motor impairments such as Amyotrophic lateral sclerosis (ALS) or stroke. However, they may also be used by able bodied individuals for as an enhanced control method for communication, entertainment etc.
A typical BCI involves four key components. These are illustrated in Figure 1.0. Firstly, signals are recorded from the brain via some neuroimaging technology. This is typically EEG but can include a host of other technologies. Secondly, some pre-processing is applied to the signals, this can include correcting for drift in the amplitude of the signals and removing artifacts (signals arising from origins other then the brain). Thirdly, features are extracted from the signals which optimally describe the components of the signal which change the most with respect to the task the BCI user is attempting. Fourth, a classification scheme is applied to classify these features into the task the user is attempting . Finally, the results of the classification are used to enact control of a computer or a device. This stage typically involves feeding the results back to the user in the form enacted control of a computer or device.
Different regions of the brain are known to form non-linear connections with one another during a variety of different cognitive tasks. This can include direct communication between different brain regions, shared input, or communication between brain regions that are part of a common pathway of information flow. Such short term connections in the brain occur during, for example, short-term memory access, language processing, motor control etc. Methods used to analyse these connections include non-linear dynamic techniques such as phase synchronization measures and complex network metrics.
Phase synchronization measures the difference between the phases of a pair of signals over a specified period of time. If the difference in phases does not change much then the signals are said to be synchronized, indicating there is likely to be some communication between the brain regions those signals are recorded from. It is important to determine whether the phase synchronization is meaningful and statistically significant and a large portion of research in such techniques involves development and choice of significance tests.
Complex network metrics measure different aspects of the network of communications between different brain regions. This can include measures such as the average number of connections to particular nodes in the network or the average path length between different nodes. The aim of these metrics if to measure how different parts of the brain communicate with one another during different cognitive tasks.
A large number of signals may be recorded from the brain. These may be described by an even larger number of features and it is not always immediately clear which feature sets are best for use in describing different cognitive events or for use in BCI control. A popular solution is to use some machine learning methods to identify which feature sets are best for discriminating different conditions.
A large variety of machine learning methods exist and new ones are always been developed. They can include wrapper based meta-heuristics such as Genetic algorithms or Particle swarm optimization. They can also include filters and transformation methods such as Independent component analysis (ICA).