Skip to main content

Important subset v0.1

Everything is over 5 random seeds
Created on September 22|Last edited on September 22

Notation

  • LM stands for label model (encoder model)
  • DM stands for the downstream (end-) model

Held-out test performance

Preds_Test report the standard deviation and mean of the predicted test labels
  • c1 means that the computation is over the proba_labels that correspond to class 1 only, i.e. proba_labels[:, 1]
  • Random encoder leaves the encoder model (~label model) fixed at random initialization
  • Hard_targets means that both LM and DM are trained on the predicted hard labels of the other model

Run set
0


Training curves as a function of gradient steps


Run set
0



Run set
0


These prediction diagnostics are for the training set


Run set
0



Run set
0



Run set
0




Run set
0



Run set
0



Run set
0