Improving the Generalisability of Brain Computer Interface Applications via Machine Learning and Search-Based Heuristics
Jason Adair · 2018
Brain Computer Interfaces (BCI) are a domain of hardware/software in which a user can interact with a machine without the need for motor activity, communicating instead via signals generated by the nervous system. These interfaces provide life-altering benefits to users, and refinement will both allow their application to a much wider variety of disabilities, and increase their practicality. The primary method of acquiring these signals is Electroencephalography (EEG). This technique is susceptible to a variety of different sources of noise, which compounds the inherent problems in BCI training data: large dimensionality, low numbers of samples, and non-stationarity between users and recording sessions. Feature Selection and Transfer Learning have been used to overcome these problems, but they fail to account for several characteristics of BCI. This thesis extends both of these approaches by the use of Search-based algorithms. Feature Selection techniques, known as Wrappers use 'black box' evaluation of feature subsets, leading to higher classification accuracies than ranking methods known as Filters. However, Wrappers are more computationally expensive, and are prone to over-fitting to training data. In this thesis, we applied Iterated Local Search (ILS) to the BCI field for the first time in literature, and demonstrated competitive results with state-of-the-art methods such as Least Absolute Shrinkage and Selection Operator and Genetic Algorithms. We then developed ILS variants with guided perturbation operators. Linkage was used to develop a multivariate metric, Intrasolution Linkage. This takes into account pair-wise dependencies of features with the label, in the context of the solution. Intrasolution Linkage was then integrated into two ILS variants. The Intrasolution Linkage Score was discovered to have a stronger correlation with the solutions predictive accuracy on unseen data than Cross Validation Error (CVE) on the training set, the typical approach to feature subset evaluation. Mutual Information was used to create Minimum Redundancy Maximum Relevance Iterated Local Search (MRMR-ILS). In this algorithm, the perturbation operator was guided using an existing Mutual Information measure, and compared with current Filter and Wrapper methods. It was found to achieve generally lower CVE rates and higher predictive accuracy on unseen data than existing algorithms. It was also noted that solutions found by the MRMR-ILS provided CVE rates that had a stronger correlation with the accuracy on unseen data than solutions found by other algorithms. We suggest that this may be due to the guided perturbation leading to solutions that are richer in Mutual Information. Feature Selection reduces computational demands and can increase the accuracy of our desired models, as evidenced in this thesis. However, limited quantities of training samples restricts these models, and greatly reduces their generalisability. For this reason, utilisation of data from a wide range of users is an ideal solution. Due to the differences in…