Consider this alternative characterization. Someone wants to fit a polynomial to some data. They pre-selected a sparse set of polynomials, which are in general ridiculously complex. Against all odds, they get a good fit to the training data. This theorem says that, because they haven’t examined lots and lots of polynomials, they definitely haven’t fallen into the trap of overfitting. Therefore, the good fit to the training data can be expected to generalize to the real data.
Shalizi is saying that this story is fine as far as it goes—it’s just not Occam’s Razor.
Good characterization. It’s worth noting that learning theory never gives any kind of guarantee that you will actually find a function that provides a good fit to the training data, it just tells you that if you do, and the function comes from a low-complexity set, it will probably give good generalization.
Consider this alternative characterization. Someone wants to fit a polynomial to some data. They pre-selected a sparse set of polynomials, which are in general ridiculously complex. Against all odds, they get a good fit to the training data. This theorem says that, because they haven’t examined lots and lots of polynomials, they definitely haven’t fallen into the trap of overfitting. Therefore, the good fit to the training data can be expected to generalize to the real data.
Shalizi is saying that this story is fine as far as it goes—it’s just not Occam’s Razor.
Good characterization. It’s worth noting that learning theory never gives any kind of guarantee that you will actually find a function that provides a good fit to the training data, it just tells you that if you do, and the function comes from a low-complexity set, it will probably give good generalization.