Project Abstract

This research aims to provide a benchmark for the classification of EEG data and to achieve high classification accuracy, driven by the increased use of EEG due to its non-invasive nature. It also seeks to overcome the high noise-to-signal ratio that plagues EEG data. The main aim of this study is to explore both traditional and modern machine learning techniques to determine whether they can accurately differentiate between experimental conditions. This approach involves preprocessing EEG epoch recordings, extracting PSD features across channels, and feeding this data into a convolutional neural network tailored for temporal-spatial signal analysis. A support vector machine classifier was also employed to evaluate whether the feature extraction methods could appropriately extract relevant information. The design includes training models on data from all participants as well as using a leave-one-subject-out approach to test whether inter-participant variability is a key factor in the difficulty of EEG classification. ICA was used to remove artifacts prior to PSD computation, ensuring cleaner input for the model. The results indicate that modern machine learning approaches, such as the convolutional neural network employed here, are more effective at extracting information from EEG data. However, inter-participant variability remains a significant challenge due to the inherently noisy nature of EEG signals. This study concludes that EEG classifiers require substantially more data to combat overfitting or must be fine-tuned to individual participants.

Keywords: Machine Learning, Convolutional Neural Networks, Classification of EEG data

 Conference Details

Session: B

Location: Sir Stanley Clarke Auditorium at 13:30 15:30

Markers: Jiaxiang Zhang, Megan Venn-Wycherley

Course: BSc Computer Science 3yr FT

Future Plans: I’m looking for work