Comment by aub3bhat | Hacker News Reader

Comment by aub3bhat | Hacker News Reader

Parent Full thread

This is a spectacularly bad advice, which "off-the-shelves" research libraries that you mention? OpenCV, LibSVM, Weka or Matlab/Octave?? Most examples/implementations in OpenCV are outpaced by Deep Learning methods.

TensorFlow, Caffe, PyTorch/Torch which today implement most of the state of the art methods were all written on average a year or two ago.

>> E.g., the crucial technical core of my startup is some original applied math I derived based on pure/applied math that's long been on the shelves of the research libraries. For the valuable work of my startup, what's in AI/ML now is in comparison at best weak, nearly silly early grade school baby talk.

Rather than vaguely mentioning libraries and startup, I suggest you offer concrete evidence behind your claims.

>> And for the education for that work, it's definitely NOT in departments of computer science. Instead look at selected programs in pure/applied math in some of the best research universities.

This is just truthiness (Math feels more "hardcore" than CS so Math department must be the better source.). Having known several people at Google/Apple/Facebook-FAIR, CS departments are typically major source of AI/ML researchers.

> ... which "off-the-shelves" research libraries that you mention? OpenCV, LibSVM, Weka or Matlab/Octave??

I think they meant "library" in the literal sense - the one with books -, and not "software libraries".

But yes, while common sense, it's empty of any meaning. Like "it's better to invest in the core math and statistical fundamentals than the specific applications". Right, but so what? Folks like DeepMind, Numenta, etc are developing new core research material, and applying in practice.

> Most examples/implementations in OpenCV are outpaced by Deep Learning methods.

Right. Except that:

- In some cases there are very good (and fast) techniques for finding out what you need quickly (like identifying faces)

- You don't want to retrain a network for your specific case (which might not have an existing model already)

- Traditional solutions might be good enough

- OpenCV is much more than just identifying images, it has several APIs like geometrical transformation, color processing, filtering, etc http://docs.opencv.org/2.4/modules/refman.html

Plenty of industrial AI is using other methods as well, it's just not where most of the current hype is, so there's a kind of bait and switch where they lead with the DL and then if you look at the products and APIs they're actually deploying and selling, the workhorses are often some mixture of very classical stats methods (like logistic regression) and general-purpose non-NN ML algorithms (like gradient boosting). One common breakdown is that the general-purpose APIs use these general methods, and then there are separate NN-based APIs for a few specific kinds of problems like object recognition in images, where NNs give a big performance increase.

Especially true for companies in the data-science niche, since DL rarely gives you much of a win for tasks like data-mining SQL databases, or when you have only modest sized data sets, but nonetheless you still need to at least offer some kind of DL solution to be perceived as a state-of-the-art AI offering. (Like with the "big data" hype wave, many companies that think they have gigantic data sets don't.) These other methods are better understood though and not in a ton of flux, so there's no need for an acquihire frenzy to get that expertise.

> This is a spectacularly bad advice, which "off-the-shelves" research libraries that you mention?

Harvard, Princeton, MIT, Stanford, Berkeley, U. Chicago, Johns Hopkins, Cal Tech, need I go on?

It's very good advice; sorry you don't understand it.

In short, the computer science people know how to write software, but for exploiting pure/applied math they don't know what software to write; that needs a mathematician, and the computer science profs rarely have enough background in math to do well with the subject.

The math is based on pure math and is heavily theorems and proofs, and about the only way to be good with that material is to be a BS, MS, and hopefully Ph.D. in pure/applied math.

The computer science AI/ML work has obtained some good results; just why the techniques work is too often still a mystery. But the emphasis on gigantic quantities of data make the work a niche and real applications rare.

> The computer science AI/ML work has obtained some good results; just why the techniques work is too often still a mystery. But the emphasis on gigantic quantities of data make the work a niche and real applications rare.

From an industrial point of view, that 'niche' includes a large proportion of the consumer internet and the 'rare' applications include online advertising, which powers two of the four most valuable companies on earth.

Reply on news.ycombinator.com