Raw EEG Data for: Learning from Label Proportions in Brain-Computer Interfaces

If you prefer to use the preprocessed and epoched data, please refer to: https://zenodo.org/record/192684 Note that this repository ontains only the visual paradigm with the N=13 subjects recorded at 31 EEG channels, as described in the above link. We copied the relevant section of the description below: This data repository contains raw EEG of an EEG experiment utilizing visual event-related potentials (ERPs) with N=13 healthy subjects. The dataset is used and described in the following journal article: <em>Hübner, D., Verhoeven, T., Schmid, K., Müller, K. R., Tangermann, M., &amp; Kindermans, P. J. (2017). Learning from label proportions in brain-computer interfaces: online unsupervised learning with guarantees. PloS one, 12(4), e0175856.</em> <strong>Please cite the above article when using the data.</strong> The data set with N=13 subjects is different to ordinary ERP datasets in the sense that the train of stimuli to spell one character (68) is divided into repetitions of two interleaved sequences with length 8 and 18, respectively. We added '#' symbols to the spelling matrix which should never be attended by the subject and hence, are non-targets by definition. The first, shorter sequence, now highlights only ordinary characters, while the second sequence also highlights '#' -- visual blank symbols. By construction, sequence 1 has a higher target ratio than sequence 2. These known, but different target and non-target proportions are then used to reconstruct the target and non-target class means. This approach which does not need explicit class labels is termed Learning from Label Proportions (LLP). It can be used to decode brain signals without prior calibration session. More details can be found in the article. In another study, the above data set was used to simulate a new unsupervised mixture approach which combines the mean estimation of the unsupervised expectation-maximization algorithm by Kindermans et al. (2012, PLoS One) with the means obtained with the LLP approach. This leads to an unsupervised solution for which the performance is as good as in the supervised scenario. Please find more details in the following article: <em>Verhoeven, T., Hübner, D., Tangermann, M., Müller, K. R., Dambre, J., &amp; Kindermans, P. J. (2017). Improving zero-training brain-computer interfaces by mixing model estimators. Journal of neural engineering, 14(3), 036021.</em> The data was recorded with BrainVision recorder. A new file was recorded for every group of 7 characters. The .eeg file contains the RAW EEG data in the format as described in the .vhdr file. Events / stimuli markers are provided in the .vmrk files. Note that there is a wrapper available to use this data in MOABB here: TODO INSERT LINK The subjects had the task to spell a specific sentence with 63 letters. In the online experiment, this was repeated 3 times and each time the online unsupervised classifier was reset at the start of the sentence.