In a metric learning system the system can be trained on known examples such that common classes (faces) are clustered together and separated from each other as much as possible. If triplet loss is used, the variance of each cluster is encouraged to be only as small as it needs to be to separate itself from other classes.
When a new face is introduced to the trained system, we can tell that it doesn't belong to any given classes because it is statistically far from any known classes - as determined by their known variance.
So now we make a new class for this face, but with only one example we don't have anything to go one to know what counts as close when checking against it in the future.
How is this bootstrapping managed?
