The recent decades have witnessed a growing scientific interest in quantitatively and automatically measuring a person's mental workload from electroencephalography data. A lot of progress has been made in identifying the temporal, spectral, and spatial features of electroencephalography data representative of changing workload and in establishing machine learning algorithms to then classify discrete levels of mental workload. However, none of the published systems has achieved a truly automated end-to-end decoding of mental workload, as they typically rely on the manual extraction of spectral features prior to classification. The current project thus asked whether convolutional neural networks (CNNs), a particular group of deep learning methods previously reported to be able to achieve spectral filtering, might allow to automate the extraction of spectral features and thus enable an end-to-end mental workload decoding framework. In order to address these questions, the feature extraction behavior of CNNs was evaluated when applied to either simulated EEG data with well known spectral compositions or real EEG data recorded during actual mental workload experiments. These studies clearly demonstrated that CNNs are indeed capable of achieving spectral filtering, and, when applied to the real EEG data, CNNs were able to reach high classification accuracies and in some cases also extracted spectral features that overlapped with features reported in the literature. All in all, the project has provided some insights into the spectral filtering behavior of CNNs and the obtained results suggest that CNNs might provide a promising route to achieve end-to-end classification of mental workload from EEG data. However, further work will be required to fully elucidate the temporal, spectral, and spatial features extracted from such data and how to improve network performance across different subjects.