I think we abstract (generalize) much slower than we can usually react with functionality. In machine learning, we actually find that abstracting IS comparable to learning,
Perhaps practicing is a form of abstraction; you practice something now to perform it later...the prepared parts of an otherwise improvised performance?
Yes I agree with your implication that improvised performance requires preparation, and interesting to think of this as abstraction. I think of it as developing a personal 'style', which I suppose is abstract.
I do like to prepare for a performance by making a new abstraction (i.e. a new combinator) but then learn how to use it during a performance. So yes, I make new abstractions in preparation offline, but I'm still working out what that abstraction 'means' online.
I think listening to music is more akin to abstraction in machine learning, as Whitehead puts it, "Art is the imposing of a pattern on experience, and our aesthetic enjoyment is recognition of the pattern." [1]
When I'm performing I think I'm listening to my own music in this machine learning sense, recognising musical abstractions which might not be explicit in the code which generates it.
Also if it's group improvisation there's learning from each other..
[1] I use this quote out a lot, but ironically it's taken from a passage that is protesting against soundbites.
Perhaps practicing is a form of abstraction; you practice something now to perform it later...the prepared parts of an otherwise improvised performance?