The human brain activity observed via scalp electrodes is influenced by the two components: volume conduction and functional connectivity of a person performing a task. When the task is a Brain-Computer Interaction (BCI) the electroencephalographic (EEG) signals represent the unique "brain print" defined by the functional connectivity that represents the interactions between different human brain zones. The other component causes trivial correlations in the observed signals and should be suppressed. Orthogonalization of signals by using autoregressive modeling minimises the unwanted conduction component, so that the residuals become correlated with the functional brain connectivity of interest. Such an orthogonalization however is limited to high-dimensional EEG data in the case of multiple scalp electrodes. The previous research has shown that the data dimensionality can be significantly reduced when a baseline in the residuals is modeled by using electrodes which make an important contribution to the outcomes. The promising results were obtained by using the Group Method of Data Handling capable of learning reliable models from multi-electrode EEG data. The project aims to demonstrate new findings on the benchmark data, capable of providing a statistically significant improvement of the BCI performance.
Research questions: (1) to extend the prior knowledge of deep neural-network models to build a near optimal model's structure on given data (2) to explore the BCI performance of conventional Machine Learning in comparison with the Deep Learning and GMDH-type neural networks.
The deadlines are as follows:
For March starters:
International applicants - 30th November 2021
UK nationals - 18th January 2022
For October starters:
International applicants - 30th June 2022
UK nationals - 5th August 2022