Statistics and machine learning for classification of emotional and semantic content of EEG

Abstract: Interpreting EEG measurements is of great relevance, both for developing underlying neuroscientific theory and improving existing applications. In this study, two networks with different approaches to time-frequency analysis and feature selection are compared on simulated and real data for semantic and emotional perception. The first network uses the Morlet wavelet transform to achieve adaptable feature selection. The second network uses a convolutional net to analyse reassigned spectrograms, in hope of improving component localisation. The result shows relatively good performance of the Morlet network with easily interpretable features, especially when combined with Grad-CAM, a method for visualising the gradients of the network to locate relevant data regions. The network using reassigned spectrograms performs less well, but comparisons with methods for ordinary spectrograms suggest that this is due to poor performance of the more traditional image-processing methods used, making it difficult to determine the effect of reassignment. Testing on novel data shows lower, but statistically significant, classification performance for emotional content, likely due both to methodological shortcomings and to the intrinsic difficulty of the problem. The study explores the use of transfer learning and finds promising results both in the accuracy on new subjects with models trained on data from others and in boosting training on single subjects by initialisation with transferred weights. Finally, the Morlet network is applied to analyse similarities between perception and memory retrieval, with significant results for networks trained on memory data and tested on perception data.