During every machine learning tutorial you'll find, there is the common "You will need to know x amount of stats before starting this tutorial". As such, using your knowledge of stats, you will learn about machine learning.
My question is whether this can be reversed. Can a computer science student learn statistics through studying machine learning algorithms? Has this been tested, at all? Are there examples where this is the case already?
I really wouldn't suggest using machine learning in order to learn statistics. The mathematics employed in machine learning is often different because there's a real emphasis on the computational algorithm. Even treatment of the same concept will be different.
A simple example of this would be to compare the treatment of linear regression between a basic statistics textbook and a machine learning textbook. Most machine learning texts give a heavy treatment to concepts like "gradient descent" and other optimzation techniques, while a statistics textbook will typically just cover ordinary least squares (if even that).
Lastly, machine learning generally doesn't cover the same material when it comes to things like model comparison, sampling, etc. So while some of the basic models are the same, the conceptual frameworks can be very different.
It's like that old joke. When asked for directions the philosopher said "Well, if I wanted to go there, I wouldn't start from here ..."
While I think each "culture" should be open to learning from the other, they have different ways of looking at the world.
I think the problem with learning statistics through studying machine learning algorithms is that, whilst ML algorithms start with statistical concepts, statistics doesn't start with algorithms, but probability models.
I really dont think so, as there are fundamental aspects in statistics that are simply overlooked in machine learning. For instance, in statistics, when fitting a model to data, the discrpeancy function that is used (e.g., G^2, RMSEA) is essential because they have different statistical properties. In machine learning, it just doesnt matter, so it is not covered at all.
Of course one could argue that you could learn those things later, but IMHO, it is better to undestand and care about some issues, and possibly in the future not care about them, then the other way around.
cheers
Depends on what you mean.
I've gotten deeper and deeper into statistics based on exposure to machine learning. (I had previously been more of a general AI guy, and hadn't had good experience with statistics, but gained greater understanding and appreciation of statistics as time has gone on.) So it's certainly a useful gateway.
I've been mostly self-taught in statistics, however, and that leaves a lot of holes. I understand and use some fairly advanced techniques, but I wouldn't have to get too far off of the path to get stymied, while someone with a firmer statistical foundation would not. And I think that would apply in your scenario of using Machine Learning as the way to learn statistics: your knowledge would be fragile.
I think that learning machine learning requires only an elementary subset of statistics; too much may be dangerous, since some intuitions are in conflict. Still, the answer to the question can it be reversed is no.
