Cost-Efficiency of Convolutional Neural Networks for High-Dimensional EEG Classification

Deep learning approaches have been at the forefront of machine learning problem-solving for the last decade. Although computationally more intensive than traditional techniques, the performance of artificial neural networks has justified their adoption for a wide array of applications. However, for small and high-dimensional datasets the large amount of learnable parameters is often a disadvantage. In this situation, the relationship between model complexity and quality gains relevance, since overfitting issues play a more central role. This is the case for Electroencephalography (EEG) classification, where it is usual to only have a small number of trials comprised of many electrode readings. In this paper, we optimize three Convolutional Neural Networks (CNNs) of different depths and evaluate them on three EEG Motor Imagery (MI) datasets in terms of classification accuracy, while also paying close attention to time consumption. The results show that the shallower ones tend to perform better at a lower cost than the deeper ones, which suggests that efforts in the direction of cost-saving may be aligned with model accuracy for small, high-dimensional datasets such as those often found in EEG.