MDS on cross-validated neural distances

Hi, I was wondering how people usually do multi-dimensional scaling on cross-validated neural distances? They are by nature unbiased and contain negative estimates because of the CV. I have a few thoughts about solutions and want to hear feedback:

  1. for each subject’s distance vector, add a common offset, which is the magnitude of lowest distance among all condition-wise pairs (usually the most negative one), which brings the whole vector >= zero. Then apply MDS on the median across subjects (because across subjects the distances are highly likely skewed considerably).
  2. just take the rank to make distance always positive.


Hi Yinan,

I would set all values smaller than 0 to 0, since your best estimate for the true value for all values smaller than 0 should be 0. :slight_smile:

Of course there might still be noise in the slightly positive values that really should be 0, but that’s just the way it is, there is no way for you to know if this is noise or a true distance.

The approaches you suggest assume that negative numbers have a meaning assigned to them, so I would be cautious.