Neural Networks for Machine Learning • Hacker News Board

Re: Neural Networks for Machine Learning

#51

Post by jberryman » Wed, Oct 03, 2012, 1:23 PM UTC

notimetorelax wrote: post #37

misiti3780 wrote: post #6

people are already complaining that you can only take the quizes once ... he had to send out an email today to everyone saying: "Many of you are unhappy with only being allowed to attempt a quiz once. Starting in week two, we have therefore decided to make up twice as many questions and to allow you to do each quiz twice if you want to. The second time you try it the questions will all be different. Your score will b…

Yeap we got spoiled with earlier classes: Algorithms by Tim Roughgarden, Machine Learning by Andrew Ng, and many more. We probably need to follow a class on gratitude. Oh well, to be fair I would donate quite a lot for each course that I enjoyed.

Actually all the entitled bitching and moaning on the ML class forum was by far the biggest turnoff of the whole experience for me. I was much happier after ignoring it and my "classmates" entirely.

Re: Neural Networks for Machine Learning

#52

Post by achompas » Wed, Oct 03, 2012, 1:29 PM UTC

rm999 wrote: post #41

azakai wrote: post #23

> Neural Networks are gradually taking over from simpler Machine Learning methods And haven't SVMs and such gradually taken over from Neural Networks?

I am not an expert in SVMs, but I consider myself fairly experienced in machine learning. In my professional experience the answer to your question is 'not quite'. SVMs have solved some problems very well, but I've had issues with them: 1. They are only for classification, not every problem is classification. The other big category is regression, for example predicting the sale price of a home rather than predicting…

SVMs are very much used for regression as well:

http://scikit-learn.org/stable/modules/svm.html#regression

Note: the scikit-learn implementation of SVMs is based on libsvm:

http://www.csie.ntu.edu.tw/~cjlin/libsvm/

Re: Neural Networks for Machine Learning

#53

Post by tomku » Wed, Oct 03, 2012, 4:42 PM UTC

notimetorelax wrote: post #38

emcl wrote: post #17

The only course that is not significantly diluted is Koller's PGM. All others have been dumbed down to a degree where they provide no challenge to the courseree at all.

It is not such a huge problem when you take several courses at once. Sadly they run them only twice a year, each time I try to follow as many as possible. I cannot follow PGM because it requires too much of my time, I'd have to abandon 2 or 3 other courses. YMMV.

I'm looking at the same problem at the moment. PGM sounds really interesting, but I think that the time investment just isn't going to be workable for me unless I drop several of my other classes. My current plan is to watch the PGM videos and try to keep up with the programming assignments as long as I can, but if it comes down to a choice of one or the other, PGM will be the one to go.

As far as "dumbing down", I've found that the Coursera classes that I've taken (Compilers, Automata Theory, Algorithms 1, SaaS and Machine Learning) have varied in difficulty quite widely. Compilers and Automata were both challenging and enjoyable, Algorithms 1 was about what I'd expect from a freshman/sophomore algorithms class and SaaS and Machine Learning were easy enough that they should be approachable to anyone with basic programming experience.

I don't feel that the difficulty in the classes that I've taken had any particular correlation with teaching effectiveness. I found Andrew Ng's ML class to be simple, but still interesting and informative - you come out of it with enough of a basic understanding to implement simple ML techniques as well as a place to start if you wish to learn more. I think that while a theory-centric class would be a nice thing to have, he's done an amazing job of making a class that can appeal to a wide range of potential students and introduce them to a field that's usually very difficult to approach.

Re: Neural Networks for Machine Learning

#54

Post by rm999 » Wed, Oct 03, 2012, 6:01 PM UTC

achompas wrote: post #52

rm999 wrote: post #41

Earlier quoted context omitted.

I am not an expert in SVMs, but I consider myself fairly experienced in machine learning. In my professional experience the answer to your question is 'not quite'. SVMs have solved some problems very well, but I've had issues with them: 1. They are only for classification, not every problem is classification. The other big category is regression, for example predicting the sale price of a home rather than predicting…

SVMs are very much used for regression as well: http://scikit-learn.org/stable/modules/svm.html#regression Note: the scikit-learn implementation of SVMs is based on libsvm: http://www.csie.ntu.edu.tw/~cjlin/libsvm/

Interesting, a quick glance at a paper on SVRs indicate they kind of work in the opposite manner of a SVM - in an SVM you try to maximize the number of points far away from the separator (taking into account class), whereas in regression you are trying to minimize this.

Do you have much background using them? I'm curious how they perform on real-world tasks.

Re: Neural Networks for Machine Learning

#55

Post by misiti3780 » Thu, Oct 04, 2012, 4:06 AM UTC

ChuckMcM wrote: post #10

misiti3780 wrote: post #7

Earlier quoted context omitted.

the only real problem with coursera is everyone is posting their solutions to github, so its gonna be impossible for them to prevent cheating. i agree with you though, that the flexibility it is offering is amazing

There is an interesting practical question here. Why cheat? If you are taking a class voluntarily over the Internet, what benefit would be gained by cheating? I presume that a large fraction of people who are doing volunteer coursework are doing it to learn, not to keep a GPA up for some other reason (sports eligibility, scholarship requirements, parental expectations, Etc.) so looking at other solutions on Github mi…

if you are never going to tell future employers, etc that you took the courses, there is no reason to cheat. if you plan on adding these things to your resume, with numbers associated with them, there is definitely an incentive to cheat. most people on the site right now are there to just learn, but coursera is hoping the ladder will eventually happen. it is a catch-22 ... sort of

Re: Neural Networks for Machine Learning

#56

Post by achompas » Fri, Oct 05, 2012, 6:56 PM UTC

rm999 wrote: post #54

achompas wrote: post #52

Earlier quoted context omitted.

SVMs are very much used for regression as well: http://scikit-learn.org/stable/modules/svm.html#regression Note: the scikit-learn implementation of SVMs is based on libsvm: http://www.csie.ntu.edu.tw/~cjlin/libsvm/

Interesting, a quick glance at a paper on SVRs indicate they kind of work in the opposite manner of a SVM - in an SVM you try to maximize the number of points far away from the separator (taking into account class), whereas in regression you are trying to minimize this. Do you have much background using them? I'm curious how they perform on real-world tasks.

Yeah, there's the SVR "pipe" concept, where you attempt to fit the margin s.t. points are close to it. It's a great alternate use of SVM's obj. function optimization.

I haven't really used SVRs aside from some exploratory work, so I can't speak too much about them. But I know they exist!

Re: Neural Networks for Machine Learning

#57

Post by switch33 » Sat, Oct 06, 2012, 6:28 PM UTC

nphrk wrote: post #49

azakai wrote: post #23

> Neural Networks are gradually taking over from simpler Machine Learning methods And haven't SVMs and such gradually taken over from Neural Networks?

Well not quite. While SVMs gained a lot of popularity for having nice properties e.g. 1) a convex problem which means a unique solution and a lot of already existing technology can be used 2) the "kernel trick" which enables us to learn in complicated spaces without computing the transformations 3) can be trained online, which makes them great for huge datasets (here the point 2) might not apply - but there exist way…

"SVMs. . .3)can be trained online, which makes them great for huge datasets (here the point 2) might not apply - but there exist ways - if someone's interested I can point out some papers)"

Please do. I want to read some about SVMs since i haven't heard that much about them.

Re: Neural Networks for Machine Learning

#58

Post by jkipp » Fri, Oct 12, 2012, 1:04 PM UTC

epsylon wrote: post #43

lathamcity wrote: post #8

I'm in the middle of the machine learning coursera course, and registered for this one as well due to interest in the material. My one complaint is that the programming assignments weren't interesting at all. The results were interesting, but the setups were mostly given to us, and we just had to code an algorithm that was in our notes. For someone who understands the basics of linear algebra and programming, it was…

Try the Learning From Data course : http://work.caltech.edu/telecourse.html A Fall run has just started (on the 2nd of oct.). It's the same version as the course given at CalTech and is more in-depth than Andrew Ng's. There is no skeleton code for the programming assignments, answers are made through quizzes. I took the summer session and learned a lot from it.

Great. Do you get a completion certificate at the end of course?