Meta learning is Learn to learn
based on tasks
(machine learning is learn to do a given task based on dataset).
It does mean that it can handle new tasks with existing status, but it is likely to learn faster and better,
because it is a good learner after learning from series of tasks.
Suppose machine learning is a function of y=f(x)
, wherein x is given input (normally data), and y is predicted output.
Then meta learning will be a function of f=F(t)
, wherein t is the input task, and f is the output function.
Naturally, meta learning is the same as machine learning.
Example 1:
Target task: Chinese to English translation. (Similar in archive of MAML)
- Support tasks: More than 20 languages to English translation.
- Query task: 2 languages to English translation
- Test task: Chinese to English translation
Example 2:
Target task: Speaker recognition
- Support tasks: Several tasks on face recognition, fingerprint recognition, MINST, ImageNet…
- Query task: Several tasks of recognition
- Test task: Vox1-H.