Isnt all learning really unsupervised at its root?

Hey all,

Not trying to trolloloolol here by any measure, serious question. I have taken some machine learning courses recently, both at brick and mortar, and of course, Andrew Ng's coursera ones, and in both we covered both supervised and unsupervised learning methods.

I guess my question is more 'meta' than a technical one, and my motivation for asking it is in starting to become more interested in how learning happens in nature, the recent thrust for more unsupervised learning ala robotics, etc.

Anyway, let us take a simple example of supervised learning - the perceptron. You have data points, and you have a label for each one, and you go about learning the best separation vector. Great. But - someone - or something - had to have labelled it in the beginning obviously. How did that person/entity label it? Clearly because they had already classified it - through their own classifier - which also means that they, at some point, learned about it, etc etc, ad infinitum.

Thus, the original 'learner' - whoever - or whatever - it was, had to have come about the labeling in an unsupervised fashion.

This means that at some point, the human had to have figured out classification in an unsupervised way. Right? Does that then not make supervised learning a subset of unsupervised learning?

To me this raises some interesting questions: For one, might it be possible, (and is there in fact some algorithm), that uses unsupervised learning to come up with labels, unto which a supervised learning algorithm then uses? Is there redundancy here, or more power? That is, can the original unsupervised learning algorithm + a later supervised algorithm be better than the original unsupervised learning algorithm to begin with?

A related question might then also be, if unsupervised learning is really the mother of all learning methods, biological or otherwise, then is supervised learning bring anything new to the table, or is it just making things easier and making use of information already given?

Put another way, and TLDR: Is there anything an unsupervised learning + supervised learning algorithm combo can do, that a pure unsupervised learning algorithm cant? ...