Why not a step function for the triplet loss?

Why not a step function for the triplet loss?

The treatment of negative vs. positive feels a bit asymmetric and biased to the positive.

That is the whole point of the Triplet Loss. It’s hard to explain everything in a post if the course was not enough. I would suggest to revisit all the material and ask a more specific question.

In simple terms, you want to model to learn when the examples are “close” (that is the reason for asymmetric treatment).

In other words, you want the difference to stand out in the positive range. I think that’s what you mean.

Thanks!