I think a CS degree is rather unique among the science degrees in that the recipients are as likely to go into Industry as they are Academia. Google, Microsoft, IBM et al. are all very hungry for CS PhDs so there is somewhat less pressure at the end of it to seek any job you can find.
Do look at Birmingham. Although I know little about the department other than that Aaron Sloman is there, that’s a good sign. At the AAAI conference this past summer, Minsky mentioned him as one of maybe three people he thought was on the right track in AI.
Whether that’s a ringing endorsement or a warning to you, you decide.
Aaron Sloman is in the field of classical AI while Domonic seems to be interested in robotics. There is generally very little overlap between the two and it’s doubtful whether that would count as a big plus.
Having said that, I’m not sure exactly how certain you are that you want to get into robotics. It might be that by the end of your final year, you want to move into something else and all my advice would be moot.
I finished my undergraduate studies in Computer Science and IT at the end of 2002, and then nearly drove myself insane trying to decide whether or not I wanted to do a PhD. After months of deliberating, I finally decided not to, and took a software engineering job.
I still don’t know if I made the right decision or not, and a few months ago I was trying to decide whether or not I should go back to uni after all. In fact, I started a thread pretty similar to yours in July here.
I’ve decided against doing the PhD a second time, mainly because I thought the PhD wouldn’t be as valuable as the experience in industry I’m gaining right now from not doing the PhD. But it’s something I still haven’t completely ruled out (even if it does get less and less likely the older I get). I still don’t know if I made the right choice; it was definitely the hardest decision I’ve ever had to make.
I’ve met a fair number of older PhD students who’ve had a number of years in industry. I think a lot were working as consultants and found the PhD helped give them an edge, but that’s only an impression. I was chatting to one recently who’s successfully combining part time study with consultancy. His motivation was partly because he felt it would give him an edge and partly curiousity about things he’d observed while working.
I’ve even met a guy who did a PhD after he’d retired. So never too late.
That’s not quite true, actually, but a rather limited view of the field. I am a PhD student in AI/robotics, and Sloman’s stuff is exactly what I’m doing. Yes, it’s not low-level control stuff, but that’s more Mech-E and EE anyway. Yes, it’s not concentrated on a subfield of AI (like computer vision or machine learning), but rather takes a broader view of the field.
That’s pretty much why I included the disclaimer about endorsement or warning. Obviously, not everyone feels that his approach is a good one.
Keep in mind that it’s not typical for one to do coursework as part of a PhD in the Australian and, presumably british system. Thus, there is much less chance to interact with the rest of the faculty and a lot less scope to participate in activities outside of your lab. My own university is home to Hugh Durrant-White, One of the two big names in robot navigation alongside Sebastian Thrun. Yet I haven’t even seen the guy around campus let alone attended any of his lectures or discussed any issues with him. In my experience at least, the British/Australian system seems to not encourage cross-disciplinary collaboration as much as the US system might.
I dunno, my experiences may be atypical though.
feel free to email me if you want to ask anything about a PhD, especially around the Sydney area. my email is in my profile.
Durrant-White and Thrun are awesome. What little I know about SLAM (and Kalman filtering) are pretty much directly attributable to them. By the by, did you see the clips of Thrun after the Grand Challenge? Tremendous. I couldn’t help but think that that was about the most recognition a CS prof is ever gonna get.
Yeah, I wonder about the differences between US and other programs and cross-discipline study. But that’s one of the reasons I mention Sloman – no matter what, as you say, your exposure to a field in grad school is going to be largely dictated by your advisor’s or lab’s interests. Personally, I see it as a problem that certain subfields are so incredibly narrow; why not choose some place with a larger view? I suppose it depends on what part of robotics is to be studied. I think of robotics as a subfield of AI, not as an end in and of itself, or just applying AI techniques to robotics. For instance, from their robotics website:
Although I should point out that it looks like funding for the CoSy project is non-existant for PhD students. (Links to other funding sources on that page.)
No, sadly I haven’t been following the Grand Challenege as closely as I should have been. Been busy getting a paper ready for CVPR (crosses fingers). But I’m downloading the video now to have a look. (addenedum: I had a look at the video and it was indeed awesome. But somehow, I’m underwhelmed by the shots taked. I’m sure they had to navigate more difficult terrain where the cameras wern’t present but cmon, all I saw was open, cleared road.)
I was actually rooting for CMU but it looks like they had a respectable showing as well. I guess you could say the real winner is robotics ;).
The problem with robotics is that it’s really hard to get a high level view of anything because the field is so immature that every time you attempt to, tiny low level details end up getting in the way as you run into unsolved problems.
“Lets assume we have a robot that knows where it is”… whats that? Well, I haven’t thought about whether to use sonar, laser or vision. What do you mean loop closing algorithm, can’t I just pick the best one? Covariance matrix? What the hell do I know about what the covariance matrix is. Yes, I assume 2D is okay for now but we’d like to move to 3D some time. What do you mean there doesn’t currently exist an algorithm for navigation in 3D (Actually, there is one algorithm, but seeing as I spent my entire honours thesis trying to duplicate his work without success, I’ve tried to expunge him from my memory :P). What are your tires made of and what’s it’s coefficient of friction? If it’s less that a certain number, you have to use a different algorithm, don’t forget, the material your driving on also affects your coeff (I’m just making crap up now)!
Sure, maybe if you had 30 frikken million dollars in funding to develop a robot like CMU did, you might be able to delegate enough lab monkeys to solve the hundreds of small problems and leave you time to look at the higher level. But most labs seem not to have such luxuries :(.
Having said that, I’m completely and utterly in awe of what Rod Brooks is doing over at the MIT AI lab. They are literally 10 - 20 years ahead of everyone else in a whole bunch of different fields. They seem to be one of the very few labs that can successfully look at incredibly high level concepts in robotics.
PS: I’m afraid I’ve been neglecting my Letterman trivia. Any hints as to your location?
In deference to the OP, I consider this not to be a hijack, but in line with the “anything to add” request, since it relates directly to the field of study.
CVPR?
Yeah, but that’s a lot of what scientists do, right? (Queue physics joke about “weightless, massless, spheres”.) The most common criticism of AI that I see bandied about is that a “new paradigm” has to be established; we don’t understand enough to even begin implementation. Hence, a high-level view that simplifies (or ignores) the details.
Now, this is not meant to take anything away from research on low-level issues or particular algorithms. They’re not only incredibly hard, but necessary, and my hat’s off to everyone making advancements in various ways. It just often seems, IMO, to be missing the forest for the trees.
I’m not as familiar as I should be Brooks’ recent work; odd that he seems to publish so little. Of course, I think they’re taking the correct approach (not like Lenat, though bless his heart if Cyc actually proves successful at some point).
That would be Indiana. Used to have it as “Axl Rose’s birthstate”. Thanks for reminding me; I change it every so often, as it gives me impetus to look into my current residence’s history.
Computer Vision and Pattern Recognition. One of the two big conferences in Computer Vision alongside ICCV.
The problem with taking the high level view is that if you seriously try and do it, you go off and do something nutty like Lenat is doing and find out that you’ve just wasted 20 years of your life. So many of the specifics of a high level view depend on getting the low level details right that most of the high level stuff people were talking about 10 years ago just looks nutty and misguided nowadays.
Check out Cog and Kismet. In particular, check out some of the proto-speech videos. Random test subjects dragged off the street would have conversations of up to 3 hours with kismet and sometimes had to be physically dragged away.
It may well be, as I know my experience has been that cross-disciplinary work is positively encouraged – I’m involved in projects not just in different areas of astrophysics, I’m also involved with collaborations with the computer science lot at Birmingham too. Or then again, maybe I’m the atypical one. 