Hello, we are interested in training a model using single-trial data (not averaged across same-image repetitions). We believe this could be helpful by functioning similar to dropout and giving us more varied samples to train the model. Do you know how we could go about obtaining that?
I think we could implement it ourselves if we had the corresponding indices used for the challenge data that denote which vertices were used; e.g., if we had the indices we could recover the same data as the challenge data via nsddata_betas/ppdata/subj0#/fsaverage/{lh,rh}.betas_session{#}.mgh I think (assuming we preprocessed it in the same manner)?
Also is the challenge space equivalent to fsaverage surface-space "nsdgeneral" ROI?
Thank you,
Paul
Hi Paul,
Yes, if you wish to train your models on single trials you can use the data in "/nsddata_betas/ppdata/subj0#/fsaverage/betas_fithrf_GLMdenoise_RR/#h.betas_session##.mgh".
Out of all fsaverage vertices, you will then want to select only the vertices used in the Challenge. For this you can use the file "algonauts_2023_data/subj0#/roi_masks/#h.all-vertices_fsaverage_space.npy" that we provided in the Challenge data release, which consists of a binary mask in fsaverage space with ones indicating the vertices used in the Challenge.
The Challenge space approximately corresponds to nsdgeneral + RSC. However, it doesn't exactly match the union of these two NSD ROIs, so to select the exact Challenge vertices you must follow the steps mentioned just above.
The Algonauts Team
Posted by: giffordale95 @ June 1, 2023, 8:21 p.m.I see, thanks so much!
Posted by: PaulScotti @ June 1, 2023, 8:39 p.m.