+ Here we are again comparing the decoded score for a true pair to a negative sample. However, rather than treating this as a binary classification task, in the equation below, we are just comparing the direct output of the decoders.
+ If the score for the “true” pair is more significant than the “negative” pair then we have a small loss. The ∆ term is called the margin, and the loss will equal 0 if the score difference is at least that large for all examples.
+ This loss is also known as the hinge loss.


Saint Olaf College

+ Cross-entropy with negative sampling
+ Max-margin loss

Loss functions

We call this a cross-entropy loss because it is derived from the standard
binary cross-entropy loss. Since we are feeding the output of the decoder to a
logistic function, we obtain normalized scores that are known as probabilities.

Learn Before

Related