I got into statistical learning theory the way a lot of people do: I trained a neural network, watched it generalize better than it had any business doing, and wanted to know why. Someone pointed me at VC dimension. I learned what it was, did the exercises, and worked out the bound for my model. The bound said the test error could be as high as several million percent. The actual test error was 3.2%.
This was puzzling, then funny, then embarrassing — embarrassing for the theory, not for me. Eventually I stopped looking at bounds at all when I wanted to predict whether a model would generalize, and that has worked out better than I expected. I want to write down why, partly because I keep meeting people who are where I was five years ago, and partly because I’m still not entirely at peace with it.
This took me a while to internalize. A VC bound (or a Rademacher bound, or PAC-Bayes, or really any classical bound) is a worst-case statement over a hypothesis class. It says: there exists some distribution and some sample where, if you got unlucky, your test error could be this bad. That’s not a prediction about your model. It’s a prediction about the class of models you’re drawing from, assuming an adversary is choosing both the data and the hypothesis.
Modern networks have hypothesis classes so large that the worst case is catastrophically bad. The actual case isn’t, because (1) SGD doesn’t draw uniformly from the class, and (2) natural data isn’t adversarial. Both facts are outside the bound’s scope. So the bound delivers what it promises — it just promises something I don’t care about.
In rough order of how often I actually use them:
None of these come with a formal guarantee. All of them are more useful than the formal guarantees I have access to.
Here’s where I’m still unsettled. Part of me says no: the theory has had forty years and still cannot produce a non-vacuous bound for a modern network. That’s a long time to be wrong about the thing you said you’d explain. Part of me says yes: PAC-Bayes-flavored bounds with data-dependent priors are creeping toward non-vacuity, and the work on implicit regularization (SGD as a stochastic differential equation, the neural tangent kernel, etc.) is producing theory that touches the actual training procedure rather than the worst case over the class.
What I’ve actually settled on is this: the bounds I learned aren’t predictive tools, they’re limits on what a worst-case argument can buy you. Once I stopped expecting predictions out of them, I stopped being disappointed, and the parts of the theory that are predictive (PAC-Bayes with the right prior, NTK in the right regime) became visible to me as a separate, more useful thread.
Why do flat minima generalize better, mechanistically? I don’t believe “flatness ≈ MDL” all the way down. Something about the loss-landscape geometry is doing real work and I don’t yet have a satisfying picture of what.