Triplet Loss

Backlinks

• [[multidimensional-mental-representations-of-natural-objects]]
  • What happens when you don’t have such information? Well then, you have to get a little creative. For those familiar with word2vec, you’ll know that a lesser-known but powerful trick they proposed is this notion of negative sampling: they found that it wasn’t enough to, during the traversal of your corpus and training your model, to push co-occurring words together, but it was also important to push words apart (i.e. randomly pick a word, and then do the opposite manoeuvre). Note the similarity to [[triplet-loss]]. In essense, we can move from pairs to triplets. But this seems like we’ve made the problem even more difficult.