cmaumet
commented
5 years ago - Think about using big voxels / lower resolution that would help with reducing dimensionalily.
Open cmaumet opened 5 years ago
cmaumet
commented
5 years ago
Discussed w/ @jbpoline:
Think whether we can incorporate constraints in the autoencoder based on the "usual" decomposition of functional maps.
How much reduction in dimension are we looking at with the autoencoder.
Supervised data: HCP, 9 pipelines from Alex's paper (maybe 3 to start with).
Compare with the average of the 3 pipelines.
JB would find how to go from pipeline 1 to pipeline 2 and then have a summary that tells you how many tools agreed on an activation in a given region.
Need to be careful with separating tasks: different tasks will have a lot of common