[gxs]

Precisely. My math courses never tested my memorization outright.

But if after pouring 40-50 hours a week into a class for a couple months didn't result in memorization of the important stuff as a byproduct, you probably weren't going to do well anyway.

I'd like to be optimistic and say that that this is how memorization began to be tested in schools. Instructors noticed that the best students seemed to have things memorizes, so they started testing this as it's an easy thing to test.

For some reason, the analogy of a doctor treating symptoms rather than the cause of the illness comes to mind.

[_delirium]

I had a statistical machine learning course whose exam was mostly factual questions, closed-notes, and oddly enough I think it was reasonably relevant, despite the fact that I usually dislike pure memorization. It didn't ask for specific formulas, but more like concepts and terminology, and how they'd be applied. It's not that these are specific things you should memorize, but that it's at least a necessary condition: if you can't, without notes, say what an expectation is, what a loss function is, what nonparametric regression is, etc., and when you might use some of these things, then you probably didn't pay attention in class or work any of the problem sets, because after a semester of actually doing the course you should definitely know all that without even really thinking.

So even if an A doesn't guarantee you actually know statistics, it's at least, imo, justifiable to say that a low grade means you definitely don't know statistics. You can always argue that you'd look things up if it was open book, but past some point if you don't know any of the material or even the basic terminology of the field, saying you could look it up amounts to saying that you could learn statistics from scratch if you needed to. It's sort of a test of, "can you hold a reasonably intelligent conversation on the topic without constantly checking Wikipedia on your smartphone for basic definitions".

(That kind of exam is probably also particularly suited to statistical ML because not knowing those things is the most common kind of real-world mistake... the details of an algorithm you can always get from an R package or Weka, but not knowing how to analyze a problem or what the main issues even are can't be solved by open-source code.)

[TeMPOraL]

> It's not that these are specific things you should memorize, but that it's at least a necessary condition: if you can't, without notes, say what an expectation is, what a loss function is, what nonparametric regression is, etc., and when you might use some of these things, then you probably didn't pay attention in class or work any of the problem sets, because after a semester of actually doing the course you should definitely know all that without even really thinking.

Terminology is easy to remember once you understand the concept, and those things you mentioned are something that you do not memorize. Those things you have to understand. You can memorize a formula, and you can memorize a list of applications of a given concept, but both of them are worthless if you don't understand on a gut level, what the concept is and thus where to apply it.

[victorhn]

Thats a good theory.

I always often hear about this kind of complains regarding algorithmic tests in interviews, that you have to "memorize" these algorithms, but i never understood this position as being constantly programming, these kind of algorithms really seem easy to do, and you don't need any memorization of them.