I've only just started looking into Machine Learning, and so far most of the examples I've seen involve starting with a training set of observations, each representing a value for a number of 'features', and using that to train a model - then evaluating the trained model against a second test set of data to allow refinements to the model. The basic operation the model is asked to do is to predict the value of one feature given others.
When talking about art (by which I would include music, poetry, etc), we can no longer assert that things are correct - but perhaps we could isolate a number of subjectively-judged features, such as whether a work is 'mysterious' or 'joyful' to a particular human judge, and then train the model to our judge's taste. However, The problem here would seem to be that having a human being add judgement features to each of our data points sounds like a very slow task, and it might often be impractical to get to the number of 'assessed works' that would allow the model to be trained.
How could the process of training a model to produce art of a certain 'taste' be done in a practicably short space of time?