3 ms·

One of the ones that prompted me to post this topic was over here, on the topic of CRF statistical models: http://www.cs.umass.edu/~mccallum/papers/ifcrf-uai200

• ewjordan 17y ago
  "Quickly grok" to "test things out" - it just isn't going to happen like that, even if you know the math behind it (FWIW, for that paper you're going to need a lot about statistics and Markov chains, but you may be able to get pretty far just by Googling the terms you don't know). Comp sci is notorious for having a huge gap between published work and working implementations. It often takes heroic efforts to put together a working version of the algorithms described, sometimes so much work that nobody has ever done it (Chazelle has some algorithms that have never been implemented in code). Even if it's plausible to do, don't be surprised if it takes weeks to get it working right. The more "on the edge" the algorithms are, the more difficult it tends to be, and it's not really likely that you'll be able to quickly experiment with new methods unless the authors have published code. This happens far less often than it should, because simple possession of a working algorithm is often enough to squeeze several more papers out of a topic, so authors have no particular incentive to share theirs, or even make it easy to create one from scratch. Computer graphics papers are particularly bad on that front, often omitting so much detail that even with significant study and effort you could never reproduce their results without researching the problem on your own anyways. Other people have offered good advice: go through it in excruciating detail, think about every equation as its own sub-problem, and make sure you understand what's happening at each step. These papers are meant for detailed study, not quick reading, even if the reader already knows a lot about the field. Google should help you through most of the math stuff, and try going to book search or the papers search if you're not finding anything online, which will give you good keywords to look for books on.
  • Torn 17y ago
    Sound (and sobering!) advice, thanks.
• caffeine 17y ago
  CRF statistical models Yeah, so, read ITILA, and do all problems up to 3, and attempt as many 4 as possible. You can skip the info. theory half and go straight to Bayesian stats and then neural nets - but you'll want to go back and learn the info theory too, eventually. BTW, you're jumping into an enormous kettle of fish called "graphical models," which are very useful but endlessly varied upon and often needlessly overcomplicated. I think what you really need is Zoubin: http://videolectures.net/mlss07_ghahramani_grafm/ http://videolectures.net/mlss07_ghahramani_grafm/ If that's a bit heavy, browse around the site for an easier introduction. I get ... how they work, but the embedded probability stuff is mostly squiggles to me Pardon me for saying so, but no, if you don't understand why each of those equations is what it is, then you don't understand how it works. That's kind of like saying "I get how the car works, but that bit in the front is just a giant hunk of metal to me." If it looks like squiggles, you need to untangle it. It's like untying a giant knot. You need to find some end, some piece of thread you can tug on, and then slowly tease each equation apart until you get how the various pieces move. (What happens if Z is bigger/smaller than what it is? Why is it there at all? What is a "clique"? Why would I add up potential functions in a clique?)
  • Torn 17y ago
    Allow me to rephrase - "I get what they (CRFs) do" then, at a high-level. Thanks for the video lecture links, I'll definitely look into those.