symbolic

This is a really useful guide on applying for the big science fellowships (NSF, NDSEG, Hertz), from a PhD student in CS at Stanford: http://www.stanford.edu/~pgbovine/fellowship-tips.htm

Microsoft Academic Search supports what nobody2008 said. http://academic.research.microsoft.com/RankList?entitytype=7&topDomainID=2&subDomainID=6&last=0&start=1&end=100 Notice that UCSD has 3.5x the number of citations as Columbia, despite only having 2x the number of publications. In the last 10 years, it has 2x the number of publications and 2x the number of citations. In the last 5 years, it has a little under 2x for both (though I think the more recent data on papers is incomplete, so it could favor either UCSD or Columbia on a per-capita basis, but UCSD still likely wins in absolute # of publications/citations). So in both absolute measures of research/impact and per-capita measures (I assume that UCSD has many more people doing ML research than Columbia), UCSD is quite a bit ahead of Columbia for ML. Interpreting this as a ranking of research for US universities: Last 5 years: UCSD #6, Columbia #14 Last 10 years: UCSD #5, Columbia #13 All years: UCSD #5, Columbia #20 So UCSD is solidly one of the best for ML research and has been for a while, Columbia is still pretty great and has steadily gotten better (or maybe others just have gotten worse). Both are awesome, exciting places to do ML but UCSD definitely has the edge. (by the way, ignore the H-index next to the institution - that's overall, for all their research, not for ML research)

Out of curiosity, I'm wondering what programs that people use, if any, to make things like more complex diagrams, flow charts, 3D visualizations, graphs/networks, etc. (the kind that can't be made easily in MS Office) for their academic papers. I've found software like OpenDX but I'm wondering what people in CS might use - I have a feeling it'll vary a lot.

But it's not just ARWU. Virtually every international ranking, whether it includes those same metrics or not, does not place 8 Asian universities in the top 15-20 or so (as is the case in some of the rankings). It quickly becomes obvious that these rankings are heavily biased toward the region's universities. There's probably a good reason that there's not much talk/discussion about these rankings on the internet; they just don't seem credible.

These rankings are really iffy: look at how many Asian universities dominate in the top 10 or top 15. I'm not saying that's impossible, but the rankings came from an Asian organization, and no other ranking (even the ARWU) seems to put that many Asian universities so high. Not just in the CS ranking, but most of their subject rankings as well. I know that Tsinghua, etc. are really great schools, but come on, they definitely don't dominate like these rankings purport.

Yeah, I'd say that's definitely true. Beyond a certain threshold, the differences in whatever measure tend to plateau pretty quickly. Or it might be just that we tend to gloss over what we don't see as "elite" and then (unwittingly) flatten true distinctions. But I like to think I'm not that snobby.

It's from 2005. I seriously doubt that much has changed, or that the few faculty hired since then drastically buck the trends shown here. While there is room for error in that data, and while the data has probably changed a bit since it was collected, it's clear that there is a trend.

Yeah, in general the NRC rankings are supposed to be more accurate than US News', but in the CS community people seem to ignore the NRC rankings. The NRC rankings in general tend not to take into account conference papers; CS profs convinced them to make an exception for CS, but they executed the exception poorly. In fact, a lot of the data across the new NRC rankings has been called out, so the NRC has finally decided to start fixing its errors... after basically telling the CS community to shut up. So they said they wouldn't fix the problems with the CS rankings, but now have decided to fix their errors since profs across many disciplines have complained. Basically, NRC shot themselves in the foot by trying to make a very complex methodology and have lost a lot of credibility since their last rankings report in 1995. Regarding faculty from top-10 schools, this is relevant: http://pages.cs.wisc...mni_matrix.html

What are you interested in? Each has its own strengths. For example, both are strong in NLP, but JHU has the advantage there. I can't speak to the other criteria of yours, but it'd probably be useful for others to know what your interests are.

Most places are going to have only the PhD in CS. Since you're interested especially in AI, I'd say definitely go for the CS program. In it, you'll do research relating to your area of interest (like AI) and you'll learn software engineering along the way, as it applies to your work. That's what I was told--it's fine if your main interest is software engineering, but if you have a strong interest elsewhere, pursue that and you'll get a good hold on software engineering anyway.

What I like about AI in particular is that, because the software and systems that you build are so complex and detailed (often extending others' systems), with so many components, researchers have often given up largely describing the implementation in their published papers--they give a quick overview, if that, and instead rely on your ability to go to their website (or their group's, or the project's) and find out the implementation details yourself. The rest focuses on the conceptual impetus for the paper, the theory, the algorithms, etc. I think that given the advent of the internet, CS is one of the disciplines uniquely positioned to split its academic work in two: the actual papers discuss more of the ideas behind the research, and the rest is left up for you to discover online. I don't know whether this is actually relevant to what you mean, since I'm not much of a systems person, but I can understand the sentiment.

True. Also one of my advisers suggested that there's more and more pressure on the 'next generation' of NLP researchers (mainly in academia, not so sure about industry research) to 'publish or perish.' Because conferences are the standard publication venue and there's so much uncharted territory in the field (given how interdisciplinary it is), it's become commonplace to see professors who publish 10+ papers a year; and the grad students with the most papers, unsurprisingly, get the best positions after graduation. But in so many of those papers, they aren't the first, or even second author; they're listed though. As self-serving as this sounds, having more people to collaborate means more papers to publish, even if your name isn't first. Which is better for your future given how much people in the NLP community publish (esp. since there are so many conferences/workshops/journals to publish in). Hahaha, that's what I call the Frida face!

Ah, the person I had talked to who suggested UMass was probably thinking of Aria Haghighi, whom she was a postdoc with. She was also suggesting to look for schools with strong machine learning faculty (something like 'a lot of NLP is about taking some hot new machine learning method and applying it to your problem'), so that's probably why she mentioned UMass as a 'top' school. If we're talking international schools, Stuttgart and Saarland both have bad-ass departments dedicated to CL. If I spoke German, they'd totally be on my list. The reason I'd want quantity over just quality (I'd choose JHU or CMU over Penn or MIT) is that when there's a lot of faculty with interests similar to yours, you have more chances for collaborations, not just with them but with their grad students. Lots of "cross-pollenation" among the different sub-areas of NLP. Of course quality is important too; I'd choose Penn or Berkeley over, say, Cornell. But all the schools mentioned with high numbers have top faculty, anyway.

Same goes for JHU, in the NLP community at least. I agree Penn is excellent, but I'm not so sure the shrinking size is because they only let in very high quality faculty--there are plenty of high quality faculty to go around (Dan Klein, for example, could have gone to Penn but joined Berkeley when there were no faculty there whose sole work was in NLP). But they've lost all their heavyweights except for Marcus, which wouldn't make Penn the best in and off itself--after all, there are people whose citation impact is greater than or equal to Marcus (Claire Cardie at Cornell, for example). Joshi is a dinosaur, Pereira is on an eternal sabbatical at Google, etc. Ani Nenkova is on her way though.

Well, my metric is more on the quality of the research, as well as the # researchers in the field--I'd say a culture of NLP is desirable for anyone interested in the field, but unsurprisingly the schools with a culture of NLP also have the best researchers. All of the schools I listed with asterisks have researchers with high impact--some of them with few people actually in NLP (MIT, Berkeley) but the people they do have are big names (Regina Barzilay, Dan Klein). I don't do anything in speech processing either; most of the work done at JHU isn't in speech though, especially now that Fred Jelinek is gone. Fun fact: three of JHU's professors in NLP (Yarowsky, Eisner, Dredze) graduated from U Penn, so you may have a point there. I'd argue it's hard to list any schools for NLP without mentioning JHU though--even so far as to say it might even be better than CMU LTI, which actually has few people solely in LTI. Most of the faculty they list are in MLD, etc. It's funny you mentioned Stanford because of the textbook--the author of it does a lot of work in speech processing.