Statistics vs. Machine Learning, fight(anyall.org)
anyall.org
Statistics vs. Machine Learning, fight
http://anyall.org/blog/2008/12/statistics-vs-machine-learning-fight/
4 comments
"What differs most is the teaching style. CS has far better lecture notes. Of course, the stats people wrote a very good book; but better lecture notes win because I can access them later and send them to people for free."
So very true. When I was a freshman in high school, we had mandatory "study hall" periods where we just had to sit in a room and be quiet. I spent most of them doing crosswords and sleeping, but I also spent a lot of time looking at printouts of these notes:
http://rutherglen.ics.mq.edu.au/wchen/lnentfolder/lnent.html
It's really important to realize that not everyone who is interested in a field has access to a library full of relevant publications.
So very true. When I was a freshman in high school, we had mandatory "study hall" periods where we just had to sit in a room and be quiet. I spent most of them doing crosswords and sleeping, but I also spent a lot of time looking at printouts of these notes:
http://rutherglen.ics.mq.edu.au/wchen/lnentfolder/lnent.html
It's really important to realize that not everyone who is interested in a field has access to a library full of relevant publications.
And AI is the red-headed stepchild caught in the crossfire, where any approach that actually produces worthwhile results is promptly excommunicated from the field.
"If I create a program which successfully predicts which humans are trustworthy and which are untrustworthy, would that be AI?"
"Are you kidding?! That would revolutionize the field! That's harder than winning the Turing Prize! That's so far ahead of anything ever done it is hard to even imagine what it would look like!"
"The algorithm exists. Its output is called a FICO score."
"Bah, that isn't AI."
"If I create a program which successfully predicts which humans are trustworthy and which are untrustworthy, would that be AI?"
"Are you kidding?! That would revolutionize the field! That's harder than winning the Turing Prize! That's so far ahead of anything ever done it is hard to even imagine what it would look like!"
"The algorithm exists. Its output is called a FICO score."
"Bah, that isn't AI."
Better examples: the face recognition in a camera, the route finding in a GPS map, and Google.
You can question if someone's models are right. You can question if their controls narrowed the experiment well enough. You can question if their interpretation makes any sense. What you can't question is if something like a t-test, LLS, or ANOVA is the best way to pull meaningful parameters out of the data. Just look at how much resistance Bayesian methods face in publication.
This is a fundamental friction that statistics has to overcome as long as its still called statistics. I think of ML as the parts of statistical research that escaped through the window opened by Shannon back when he invented information theory. It's a bird now, free to invent its words for the world and try crazy stuff that the religion of "Statistics" could never accept.
This isn't to say that Statistics isn't growing. The article itself does a good job pointing out just how similar recent Stat has been to ML. However, if you see a research paper in some of the more core, less data intensive sciences that dares to drop "SVM", "Ridge Regression", "Clustering", "Bayes", or god forbid "Machine Learning" itself you see scowls: isn't your data normal enough? Why do you need to do something fancy when I can work out the z-score of that result right here on my pocket calculator/slide rule.
(Lets go ahead and concede that ML certainly has a lot of broken yet overhyped parts which helps form the nucleus of an argument not to infect scientific knowledge with some untested infrastructure. Growing pains.)
It's a classic fight between tradition and innovation with all the usual arguments available, but what makes this different is that such a huge community of people who thrive off the image of really, really knowing things pretty much take the frequentist methodologies as an unimpeachable gold standard. Things can get dicey when you start to ask what the actual meaning of a p-value is, how we really know anything about "estimators", why people work so hard not to use computers. It works! Stop asking questions.