most people aren’t defining their subspaces. Take the issue of PCA. If experiment cond A pushes a latent to have higher variance than condition B, comparing the two artificially thru alignment artificially inflates “orthogonaliy”. Better to define geometry based on variables and decompose.
Reposted from
Grace Lindsay
Diversity of responses to this also makes me curious: what do you assume a paper is saying if it says a neural population is using 'orthogonal subspaces'? Or what analyses would you say are needed to make this claim?
Comments