Python: Advanced Guide to Artificial Intelligence

Triplet loss

The triplet loss learns the score vectors for the images. The score vectors of face descriptors can be used to verify the faces in Euclidean space. The triplet loss is similar to metric learning in the sense of learning a projection so that the inputs can be distinguished. These projections or descriptors or score vectors are a compact representation, hence can be considered as a dimensionality reduction technique. A triplet consists of an anchor, and positive and negative faces. An anchor can be any face, and positive faces are the images of the same person. The negative image may come from another person. It's obvious that there will be a lot of negative faces for a given anchor.

By selecting negatives that are currently closer ...