Estimating mental workload from brain signals such as Electroencephalography (EEG) has proven very promising in multiple Human-Computer Interaction (HCI) applications , e.g., to design games or educational applications with adaptive difficulty, or to assess how cognitively difficult to use an interface can be. However, current EEG-based workload estimation may not be robust enough for some practical applications. Indeed, the currently obtained work-load classification accuracies are relatively low, making the resulting estimations not fully trustable. This paper thus studies promising modern machine learning algorithms, including Riemannian geometry-based methods and Deep Learning, to estimate workload from EEG signals. We study them with both user-specific and user-independent calibration , to go towards calibration-free systems. Our results suggested that a shallow Convolutional Neural Network obtained the best performance in both conditions, outperform-ing state-of-the-art methods on the used data sets. This suggests that Deep Learning can bring new possibilities in HCI.