## Origin of data ## The data were recorded at NeuroSpin, Saclay (France). *Contributors* * Florent Meyniel (principal investigator) * Micha Heilbron (data collection) * Maxime Maheu (data collection) * Sebastien Marti (technical assistance, director of MEG facility). *Full postal address* NeuroSpin Commissariat à l’énergie atomique et aux énergies alternatives Centre de Saclay Bâtiment 145 – P.C. 156 91191 Gif-sur-Yvette Cedex France *Please cite the paper if you reuse those data* Meyniel F (2020) Brain dynamics for confidence-weighted learning. Plos Computational Biology (accepted) ## Organisation of files ## Each subject is uniquely identified by an integer that appears in the file names. *Demongraphic data* Basic details: recording data, age and gender of each subject. Data collection for subject 18 was interrupted (technical problem). The subject came back another day for a new and successful recording (registered as subject 25). There are thus no data for subject 18. *MEG DATA* The raw data are provided in the constructor's format, fif, which can be read by most analysis software (e.g. [FieldTrip][1], [MNE][2]). The acquisition was split into 4 sessions. Those data have been processed with the constructor software MaxFilter (see paper for details). Subject 11 has no session 1 due to a technical problem in saving the data. *Pupil data* Due to technical problems, the data was lost (not saved) for some subjects, inserted as an external analog signal in MEG recordings, or as an EyeLink log file. In order to facilitate the use of the data, we provide under the same format (matlab's .mat) the available data, after preprocessing (see the paper for details). For each subject, the stimulus-locked, epoched data is provided as a set of two matrices and one vector: - *alltrials* is a matrix (number of trials x samples within the epoch) of preprocessed but not baseline corrected data. Trials with NaN correspond to excluded trials. - *timevec* is the peri-stimulus time vector - *trialinfo* can be ignored (the order or rows in *alltrials* corresponds to the order of trials in the experiment). *Sequence of observation and behavioral data* For each subject we provide a CSV file with the following columns: - Session number -> 4 sessions in total - Trial number -> within a session - Sequence -> observed sequence of items, coded with 1s and 2s - Generative p(1|2) -> actual generative transition probability - Generative p(2|1) -> actual generative transition probability - Is question trial -> if true, then two questions were asked after the stimulus presented on this trial - Subject est. p(1) -> question #1: probability estimate of the next item being 1 - Subject confidence -> question #2: confidence about the probability estimate - RT (s) prob. est. -> RT on question #1 - RT (s) confidence -> RT on question #2 NB: in trials without question (which is the majority of trials), the 4 last column have the placeholder value 0. *Scripts for preprocessing* This dataset contains raw data that will need to be preprocessed in order to be analyzed. A key aspect of the preprocessing is to read the recordings with the associated events (which stimulus was presented, when, etc.). The events are coded in the MEG dataset as a combination of signals on channels collecting triggers from the stimulation computer. We provide two scripts, that were part of our preprocessing pipeline (with FieldTrip), to epoch the data around each stimulus onset (**FieltTrip_DefineTrials_NoQuestion.m**) and each question onset (**FieltTrip_DefineTrials_Question.m**), together with various pieces of information (trial number, stimulus type, etc.). Users interested in analyzing the data with FieldTrip can use those scripts directly. People using another software can have a look at those codes in order to know how to read the triggers and collect events in the dataset. The data analysis leverages the use an ideal observer model (see below). We also provide an example script to use the ideal observer and extract key variables (e.g. surprise, confidence about the prediction) in **example_ideal_observer.m** ## Ideal Observer inference ## In the paper, the MEG signals on each trial are compared to different aspects of an optimal inference, in particular, how surprising is the current observation given the previous ones, and how confident one could be about the prediction of the current observation.\ Such predictions depend on assumptions: how the observer believes that the observations are generated. In the paper, we take the perspective of an ideal observer, whose assumptions about observations correspond to the true generative process (described in the paper). This observer is "ideal", because, given the observations received, it makes the best (Bayes-optimal) estimation of the probability of the next observation in the sequence. This ideal observer model was introduced in this paper:\ [Meyniel, Maheu & Dehaene (2016) Human Inferences about Sequences: A Minimal Transition Probability Model *Plos Computational Biology*][3] The corresponding matlab code (used for this project) is available at: https://github.com/florentmeyniel/MinimalTransitionProbsModel More recently, we also made a Python version of this ideal observer (which we made more general that the matlab code):\ https://github.com/florentmeyniel/TransitionProbModel ## Data collection ## The MEG data were collected with an Elekta Neuromag machine. The machine has: - 102 MEG magnetometers - 204 MEG gradiometers MEG Acquisition parameters: - Acquisition frequency: 1000 Hz - Online high-pass filter: 0.10 Hz - Online low-pass filter: 330 Hz Pupil size was recorded with an EyeLink 1000 eyetracker. [1]: http://www.fieldtriptoolbox.org/ [2]: https://mne.tools/stable/index.html [3]: https://doi.org/10.1371/journal.pcbi.1005260 "PCB 2016 paper"