This project was done by :
Heejung Jung (Graduate Student)
Xiaochun Han (Post-Doc)
Deepanshi Shokeen (Graduate Student)
Project: MVPC – classify behaving animal data in searchlights using SVM
- P1. 1 out of 20 conditions
- P2. Classify taxa (1 out of 5), training on videos with 3 behaviors and testing on videos with the left-out behavior.
- P3. Classify behaviors (1 out of 4), training on videos with 4 taxa and testing on videos with left-out taxonomic category
Parameters:
- The 3 analyses (P1, P2, P3) were conducted separately for the attention and behavioral task.
- search light radii: 10
Analysis involved
-
GLM: modeled 20 separate regressors with nuisance regressors derived from fmriprep
-
searchlight: using a 10 radius, pattern classification was performed using a linear SVM classifier within a surface-based searchlight, conducted per participant.
-
group average searchlight one-sample t-test & two-sample t-test
- one-sample t-test
- Group average searchlight maps were compared against the chance performance
- chance performance was 0.05 for P1, 0.2 for P2, 0.25 for P3
- two-sample t-test
- For example, for the behavior classification for beh attention > tax attention task
- We hypothesized that the behavioral classification performance would be higher for the behavior-attention task, as opposed to the taxonomy-attention task.
*Note that we used a one-sided uncorrected p-value of 0.05.
- visualization of group average searchlight one-sample t-test & two-sample t-test
Explanation about the 1 - cross validation
- We realized in that in our code, we simply ran cv = mv.CrossValidation(clf, mv.NFoldPartitioner(attr=chunks)).
- After realizing that the default was mv.mean_mismatch_error we subtracted the searchlight outputs from 1, in order to account for the fact that we calculated the average mismatch error.