What is metric learning in kaggle competitions?

Going through some kaggle discussions I came across the term Metric Learning. And I am not able to understand what it is and how to implement it?

Things I understand. Metric Learning is useful when we some classes with few examples in the training set. Metric learning tries to find some distance between our test image and input image (I guess).

Can someone please explain this or give pointers on where to read about it?

Did you ever find an answer to this?

I was thinking that maybe it means the metric or loss used during training also has parameters that can be optimised for.