Stanford CS330_ Deep Multi-task & Met...

Learn the embedding in a parametric way and the remaining is non-parametric, like the nearest neighbor.

If there is more than one shot, need to compare the prototypical example from each class.

Original comparison is hard to deal with, that is way the soft version prediction is introduced.

Consistency: the learning procedure will improve monotonically with more data. It implies generalization but not vice versa. black box method is not consistent.