Answer by Stephen Miller:
Inherent biases: I did pretty heavy undergrad research at Berkeley and am a PhD student at Stanford. While I have plenty of friends and conference buddies at MIT and CMU, and visited both before making my grad school decision, I'm not nearly as knowledgable about their strengths and weaknesses. It's far more about surface level impressions from visits and conferences. Also I know next to nothing about HCI, so I won't try to touch on that. Whereas I work in Robotics, so I'll be skewing this inordinately towards related disciplines to that. If you want to know exactly how I am biased, Google stalk to your heart's content.
Obvious disclaimers: These are the top four universities in the world for AI, and I personally found it extraordinarily difficult to choose which I wanted to go to for grad school. Any strengths or weaknesses are being graded on a very, very steep curve: they're all fantastic for whatever you're interested in. If I mention or fail to mention any professors or groups, it's only because I'm personally more or less familiar with their work; it's not a quality assessment at all. I'll probably also forget groups which I really shouldn't be forgetting, and use colloquial names which seem disrespectful. None of this is intentional. I'll also use distinctions like "pioneering/fundamental" vs "young/rockstar/rising", which simply means that the former has made contributions to their field which are time-tested and game-changing. The latter might be equally good or better, but they haven't been around long enough to talk about their "legacy".
Fields I'll mention: I'm only bringing up things which are widely considered "AI". This includes Machine Learning / Graphical Models, Planning, Computer Vision, Robotics, Comp Bio, NLP, and a small bit of Graphics as it relates to the above fields (e.g. I'm not worried about things like raytracing or simulation, mainly because I know very little about them.)
In decreasing order of familiarity:
Stanford could be characterized by having an inordinately high percentage of big names who have made fundamental contributions to their field, but with fewer groups working on elaborate demos / end-to-end software systems. You'll find groups doing pure Machine Learning (Andrew Ng in Deep Learning, Daphne Koller in Graphical Models, others in the stats department), where the particular application is less important than the underlying mechanism. Jure Leskovec applies learning and graph-theoretic techniques to model and analyze social networks. Pure Computer Vision is steadily growing, with (in addition to being the ML groups' pet application) Fei-Fei Li and Silvio Savarese being very well respected in the field; there tends to be an emphasis on the high-level here, with more object/scene/action recognition than, say, stereo, edge detection, segmentation, etc. Those roles tend to be filled by the Robotics and Graphics groups. Robotics tends to be split between perception (Sebastian Thrun's group), control/haptics/feedback (Oussama Khatib, Ken Salisbury) without much bridging the gap. Allison Okamura's lab is big in the Surgical Robotics space. The Graphics department has increasingly started working in 3D perception (an obvious strong suit after years of working with 3D models), usually with an emphasis on geometric reconstruction/registration — Vladlen Koltun, Leo Guibas, and Marc Levoy are coming to mind. NLP is very strong, with Dan Jurafsky, Chris Manning, and Percy Liang comprising a pretty sizable fraction of the field. Finally, Comp Bio is massive here, with Serafim Batzoglou, Gil Bejerano, and Daphne Koller making up the AI front.
Brief: a high number of extremely "core" people in their respective fields, relatively few of whom are currently young and tenure-seeking. Fantastic research in particular subjects listed above, with a bit less of an internally collaborative spirit (though there are always exceptions, and cross-department collaborations are particularly common). If you want an analogy, I think comparing it with a hugely successful Sillicon Valley research company — massive prestige for all the higher ups, fun atmosphere for workers, lots of exports and a well-defined brand name — would be fair.
Berkeley is a very diverse group with, from what I can see, a high degree of collaboration. The ratio of legacy/pioneering professors to younger/multi-faceted ones is lower, which is in no way a value statement for or against it. Where at Stanford everyone is a huge leader in a preexisting field, at Berkeley you'll find more that seem willing to try new things. Of course, there are pioneering researchers here as well. On the Machine Learning front, Mike Jordan and Martin Wainwright are a pretty hard-to-beat team in the stats fundamentals. There's also core AI/Logic research being done by Stuart Russell, which, as far as I can tell, doesn't have an analog at Stanford. Computer Vision is extremely strong with Jitendra Malik, (new hire from CMU) Alyosha Efros, and Trevor Darrel all very high profile researchers with independent labs. While the high level (e.g. object detection) is certainly there, it's probably fair to say that there's a bigger diversity of lower-level tasks: 3D from a single image, segmentation, rethinking classification (depending on what aspects of Prof. Efros' research transfers), etc. On the robotics front, Pieter Abbeel is a younger professor with a wide range of interests, from reinforcement learning to control to perception, and a big emphasis on building end-to-end systems; Ken Goldberg is big on cloud and surgical robotics, as well as grasping work and tons of collaborative efforts with media groups; Claire Tomlin works on control and planning, particularly for hybrid systems; Ron Fearing does mechanical design and control, typically for small biomimiteic robots. The AI side of Graphics I'm fairly unfamiliar with here, though James O'Brien has been known to collaborate in simulation-for-robotics projects. NLP is also very strong, with Dan Klein being a veritable badass.
Brief: a diverse group who seem a bit more interested in mixing things up, with fewer "fundamental" names but a higher energy level. There's little overlap between labs' expertise; hence the collaborations. If Stanford is a prestigious SV company, Berkeley is the lean startup counterpart.
(Here's where my personal relationships break down, so the number of specific names I give is probably going to decrease exponentially. Someone else should jump in who knows these schools better, because there's no way I'm giving them their due.)
CMU is the epitome of diversity — think Berkeley with the number of professors multiplied by 50…OK, that was a slight exaggeration. But everything is going on there; the joke is if there's something you want to do, there's probably a department named after it. Case in point, there's an entire Machine Learning department with a separate admissions process. From there I recall a huge emphasis on solving big graphical models applied to particular domains (notably in the bio and health regimes). With that is an emphasis on big data ML — Carlos Guestrin is the primary name that comes to mind, though he recently moved to UW. There's also a lot of more fundamental algorithmic work, e.g. the Professors Blum, and IIRC an emphasis on life-long learning. The Vision groups tend to have close ties to the Robotics community, and many hold dual-appointments with the Robotics Institute. Martial Hebert and (til very recently) Alyosha Efros frequently collaborated, working on 3D, segmentation, object detection, and other similar things. In the pure Robotics side, Drew Bagnell is huge in learning, control, and plannin; Sidd Srinivasa does perception, grasping, HRI, and many others I'll feel guilty if I forget to mention; Matt Mason is widely known for doing fundamental work in grasping; Manuela Veloso works on perception and planning for mobile robots + robot soccer; and Chris Atkeson is big on agile control. I know absolutely nothing about Graphics, Comp Bio, or NLP here, though I know that there is an entire NLP department. So my lack of knowledge shouldn't remotely indicate a lack of research going on.
Brief: I always got the sense that this was an extremely community-oriented atmosphere, with particularly large emphases on end-to-end systems and big data. With an entire Robotics Institute, there's a lot of work done on making functional systems, and you're far more likely to see robots in action here than anywhere else. In this respect, I'd say the research is much more application oriented than some of the "pure science" work of the above groups; though in my view that's a strength, not a weakness. If Stanford is a powerful tech corporation and Berkeley a lean startup, CMU is a giant think tank. There's no well-defined brand here, but there are tons of experts working in overlapping fields, trying to make a well-defined goal happen.
MIT will forever be synonymous with AI thanks to Marvin Minsky. I think it's fair to say that his legacy still hangs there pretty strongly. There's a lot of work being done here in the more fundamentals of classic AI: for instance, belief space planning with Leslie Kaelbling and Tomas Lozano-Perez and very cool cognitive investigation work in Josh Tenenbaum's group. Tommi Jakkola is big in Machine Learning, with a primary application in genetics. On the Computer Vision side, Antonio Torralba is a big younger name in the field, doing object and scene recognition but with a knack for ingenuitive side projects; Bill Freeman is more on the pioneering side, with major contributions to computational photography. Aside from frequent application work in the above AI groups, robotics also has a big Aero-Astro component — Nick Roy and (recently) Julie Shah both being big names in aerial robotics. Otherwise, Russ Tedrake is big on optimal control for legged robots (with incredible demos to show for it), Daniella Rus works on distributed robotics, Seth Teller does HRI (a broad generalization — there's a lot of variety in his lab). On the perception side of Robotics, John Leonard is very big in the Mapping and Navigation (e.g. SLAM) community. I know very little about the broader Graphics community, but Fredo Durand is a big name in, among other things, computational photography. I also know embarrassingly little about NLP, but it's certainly there. Comp Bio is also huge here, though I'm not familiar enough with the labs to go into detail. Finally, collaborations with the Media Lab are pretty common, so a lot of AI-for-art-installations awesomeness is going on there.
Brief: MIT is sort of an outlier in all respects. It has many people who have done pioneering work in the field of AI, and if you're looking for "fundamental AI" work (especially planning) it's arguably at the forefront. Their robotics groups seem to excel in navigation, control, etc (what I'd call low level) and abstract reasoning (what I'd call very high-level), without as big an emphasis (that I've seen) on semantic perception. And there's a big Comp Bio group. They're probably more like Stanford in the research style, but with more of the Berkeley breakdown of topics — hard to communicate what I'm going for here, I know. I don't think I have an analogy for MIT to go with the others: there's nothing quite like it. They basically started this field, and they remain an entirely unique thing.
And now that I've probably misrepresented and pissed off a variety of groups, and almost certainly forgotten about incredible professors, it's time to run away.