As I was looking at yesterday’s post, I started thinking about a sort of Turing Test for fonts: Would it be easy or hard to design a randomized font — in the style of the one I showed yesterday — so people would not be able to tell that the randomness was machine generated?
And then I realized that it would be quite easy: You could use a “big data” approach. First analyze a lot of samples of actual human writing, then use those to train a machine learning algorithm. You can then use that algorithm to generate new writing samples. It’s one of those problems that is actually quite amenable to a “big data” machine learning approach.
But then I started thinking, could we start to arrange all human abilities on a scale from “easily faked by big data” to “not at all fake-able by big data”?
Some things, like generating randomized fonts, are on the easy end of the spectrum. Other things, like maintaining a long term intimate relationship, are probably way off on the difficult end of the spectrum (or at least, I’d like to think so).
But what about everything in between? Driving a car has turned out to be more tractable than people had once thought, as have chess and rudimentary translation between natural languages.
I wonder, is there some litmus test we can apply, to get a rough sense of how easy or difficult it would be to emulate any human task via machine learning, given sufficient data showing humans themselves doing it?
This is an interesting take on machine learning of fonts that I just saw today:
http://erikbern.com/2016/01/21/analyzing-50k-fonts-using-deep-neural-networks/
I have an idea for how you might estimate learnability. Machine learning can be thought of as a kind of interpolation technique. If you can find a way to represent the data so that a weighted average of a few nearby examples is a good approximation of the correct answer, then it will work pretty well. But with a relationship, the dimensionality of the space of inputs is so large (it includes all the previous sentences said in the relationship) and the model to be learned is so complex (as complex as a human personality) that finding enough examples to have a few nearby inputs (full relationship histories) to a given input will be impossible.
A new “uncanny valley.”