Validation of hyperparameters

[SophieHerbst]

I started using the multivariate rank-1 CSC, and am unsure about how to select the parameters, e.g. the regularization and number of atoms. Which output should I look at to assess the fit independently?

Also, I would like to chose a longer time per atom and not apply a high-pass filter > 0.5 Hz to be able to look at slow dynamics. Do you have any advice the parameters regarding this?

[agramfort]

I started using the multivariate rank-1 CSC, and am unsure about how to select the parameters, e.g. the regularization and number of atoms. Which output should I look at to assess the fit independently?

number of atoms it's like choose the number of components in ICA. I would select more rather than too little. Regarding reg param I think we have a good heuristic by default.

Also, I would like to chose a longer time per atom and not apply a high-pass filter > 0.5 Hz to be able to look at slow dynamics. Do you have any advice the parameters regarding this?

slow atoms will require long atoms and will slow down algo. Can you then resample to reduce data size? I would say use atom which are 1.5x the duration of a pattern you expect to see. If you look for patterns of 400ms than use 600.

[SophieHerbst]

Thanks! Computation time is indeed an issue, and seems to increase with more atoms asked for and smaller reg.

Regarding reg param I think we have a good heuristic by default.

It is 0.1/ 0.2 depending on the examples, what do you mean by heuristic?

[jasmainak]

Remember to standardize your data -- otherwise the scaling of your reg parameter may not be good.