What would be the loss function of the relation network? We will use the Mean Squared Error (MSE) as a loss function. Although it is a classification problem and MSE is not a standard measure for classification problems, the author of the relation network depicts that, since we are predicting relation scores, it can be considered a regression problem; despite that, for ground truths, we can only automatically generate {0, 1} targets.
So, our loss function can be represented as follows:
Where are the parameters of our embedding function, , and relation function, respectively.