hi,
thanks for the input.
- What do you mean by the “scale” parameter? The
std? - Yes, you are right. and for the MEG, I actually have a noise covariance matrix that, so i know how they covary/correlate. But i have no idea, how to model that and/or use that strong prior information. do you have any idea?
- Actually, using variational inference with minibatches is the way I do it now. It decreases the computation time from ~20 hours to 20 minutes.
Concerning your question regarding the shape of the matrix: EEG or MEG data is sampled at a regular interval like 1000Hz. It is the strength of these methods that they allow you to infer when an effect happens. So, let’s say, I want to analyze the cortical response to an auditory (e.g., a sound) compared to a visual (an image, for instance) stimulus. I would present the audio ~100 times and the visual ~100 times as well. I have triggers in my data, so I know exactly, when this stimulation happened. I would then cut out the data around these triggers (like 300ms before the stimulation to 700ms after). Because of the very low signal to noise ratio, I would then average the 100 visual and 100 auditory trials within each participant. Assuming a sampling rate of 1000Hz, I would have 1000 samples per channel for each condition and participant.
This leaves me with a (n_channels, n_samples) matrix per participant and condition, which reduces to just one of these matrices per participant before I apply the statistics, because I am interested at what channels (where?) and what samples (when?) there is a difference between the two conditions.
So, in the end, I concatenate all these matrices for all participants, which leaves me with said (n_participants, n_channels, n_samples) matrix.
Does this make it clearer?