A quick idea.
I think I have a start of an idea to meassure data. If you have a simple machine learning model.
Feed that model, y = model(np.random.rand(1)) and meassure the time for the total iteration to converge. To reach the data.
This time value is then some measurement of the complexity of the data.
One thing that could be of value from this is that an ordinary data like sin() needs a large amount of initial iterations. I think this could be compensated like this in other problems.
// Per Lindholm 2018-05-02