simonucl / hjcl Goto Github PK

Hello, I have recently read your paper and I find it very informative. But I am a little confused about the paper. In Instance-level Contrastive Learning, are Xi_l and Xj_l positive pairs in their subsets(above level l) and are they multiplied to get the cosine similarity? And in Label-level Contrastive Learning, i just don't know what the metric mean. Are Yi and Yj all labels ture(1) or false(0) correspond to specific instance i and j, if so how does the distance of News and Classfields figure out to be 4 and the distance between United Kingdom and France is 1?

Best whishes!