Why is it called machine learning while it is actually normal statistics and control theory?
There isn't anything in there that learns, there is only a data model and a strict algorithm.
It should rather be called Data Mining.
There isn't anything in there that learns, there is only a data model and a strict algorithm.
It should rather be called Data Mining.
Machine learning is kind of like statistics for street-fighting. It doesn't care about Bayesian v. frequentist philosophical debates, resorts to cheap tricks like regularization, dropout, using many dumb methods that work better than trying to find the smartest method (ensembles and funny tricks like random forest). If it works in a well-designed test, use it.
There is more than one path to enlightenment. One man's 'data mining' is another's reasoned empiricism, letting the data do the talking instead of leaning on ontological preconceptions of how data is supposed to behave.
Machine learning often works well but you don’t always know why. Stats doesn’t always work and you know precisely why not. (because your data doesn't fit the assumptions of OLS etc.)
> There both about inference
This is actually incorrect -- a large swath of machine learning algorithms (like deep learning and SVMs) are unconcerned with the underlying distribution of your data, and are perfectly happy to generate predictions using a completely uninterpretable model.
> Statistics [...] comes out of math, has a greater emphasis on proofs
There is an entire branch of machine learning, known alternatively as learning theory or complexity theory, devoted to the analysis of algorithmic generalization error. Generally speaking, any popular machine learning algorithm has a proof of its (1) performance and (2) ability to generalize.
> It doesn't care about Bayesian v. frequentist philosophical debates
If this is true, it's only because machine learning has firmly parked itself in the Bayesian camp wherever possible. Naive Bayes, regularization, dimensionality reduction, and graphical models are either motivated by or easily interpreted as Bayesian techniques.
perhaps I used the term incorrectly but a dictionary defines inference as "a conclusion reached on the basis of evidence and reasoning."
bottom line... similar objectives, slightly different if often overlapping emphases and bags of tricks.
Machine learning practitioners care more about computational performance or generalizability beyond seen data than they care about properties of a data-generating distribution.
Your "bottom line" summary and repeated reference to algorithms as "bags of tricks" are reductionist statements suggesting you don't understand the nuanced, but important, differences between the fields.
I had a distinguished math professor who taught me that mathematics is the application of cheap tricks and bad jokes. He didn't take himself too seriously. When I learned about random forests I couldn't help having a good laugh, it's a cheap trick in the finest sense of the word.
Regularization is not a 'cheap trick'. It is an essential technique for mathematically ill-posed problems, where no unique solution exists. This is extensively studied and well-understood, and is used in many fields outside of machine learning.
So how is it different from statistics. It isn't, but usually the focus in ml is on big data and the accuracy of the prediction, while statistics tend to focus more on the explanation of the data.
I believe the reason for the difference is one sprang out of computer science and revived research into AI and therefore is more hip. The other didn't and is still fighting with its image of being dry and boring.
Where the structure of such model depends on training data and is not chosen by the programmer ahead of time.
(Manually structured NNs, RBMs and HMMs do not really count.)
A lot of stuff is the same between machine learning and statistics with different terminology. With machine learning often more focussing on the practical side rather the theoretical side (or the real probabilities).
It's good to cross pollinate and learn from each other. Such as the machine learning field learning to create better models with Bayesian statistics, and statisticians learning from models like support vector machines.
I also look forward to more mathematicians and statisticians exploring deep learning, and finding more rigorous scientific explanations why they work so well (and this can help us making them even better).
(I come from an Econometrics background, but wrote my master these about machine learning).
That said, Machine Learning is different from statistics. Statistics are mostly concerned with modeling real world phenomena. Machine Learning is trying to build models that can take a training set and generalize to unseen data without building a specific model for that dataset.
In this light, most Machine Learning is just a rebadging of Classification theory that sounds cooler.
Somewhere in the future, ML will fall off the hype curve [1] and something new will come along for boring old disciplines to rebadge themselves as. The good news is that these disciplines all leave bits of themselves behind and the things they were pretending to be become better and better defined. It's amazing how few job posting for Data Scientist require a PhD a Stats these days, but it was all the rage for a year or two. Now a reasonable Data Scientist can train up on Corsera in a few weeks.
1 - http://4.bp.blogspot.com/-eL79PoJLFVY/UfSulEQrdfI/AAAAAAAAAw...