Open changhoonhahn opened 5 years ago
We had a brief discussion about this before. Essentially the task is to reconstruct the data vectors from the lossy compressed representation and use a metric (say, mean sq error) to quantify it, right?
We are comparing the final posterior distributions of course, but there's no ground truth there.
Compare using Gaussian pseudo-likelihood the posterior derived from score, PCA, etc compressed data vectors to the posterior from the the peak count data vector to validate the different compression methods.