robot_control

Bumping thread. D-date near. Appreciate any comments and advice. Thanks!

Hi, I have been fortunate to get two wonderful offers: Berkeley and UW. I would appreciate any comments about them and advice on where to go. I visited both places, and liked both universities. Both cities seem to be huge tech hubs (Berkeley more so than Seattle, but the latter is great too), have good weather, and look gorgeous. I can see myself doing good work at either place and also living comfortably. For many people, Berkeley would almost look like an obvious choice, but surprisingly I find myself leaning towards UW. I have summarized my reasoning below. Am I shooting myself in the foot? Berkeley has many great researchers in my area, but they are not as accessible as I had imagined. Most top POIs are oversubscribed with 15+ grad students and postdocs. Hence, it would be hard to get face time with them, or I have to compromise and work with junior faculty. On the other hand, UW CSE is expanding at a good pace with great funding. Many top faculty were hired over the past 2-3 years and are in the process of expanding their labs. Sergey Levine joined recently, and the general consensus is that he was the brain behind most of the robotics work at Berkeley over the past 2-3 years. The situation boils down to sub-optimal POI support (either face time or researcher) at a more prestigious place vs ability to work freely with anyone I want at a (slightly?) less prestigious place. Can anyone offer advice on how to navigate this landscape? Is the difference between Berkeley and UW very significant, that too in ML/AI/robotics? I believe both are top 5 schools. Thanks!

Got into the EE program at UW today; still waiting to hear from CS. I'd much prefer the latter. @emmm What's your research area? Computational biology?

@ev a. Sounds very similar to my own story However, I knew from my sophomore year that I wanted to do robotics. Only later did I figure out that CS builds the brain for the robot whereas EE is more like the central nervous system. I associate mostly with the ICRA and IROS communities (and maybe CVPR) which do have decent representation from EE people. So I don't actively abhor EE, but given a choice will obviously pick CS. I however agree that much of EE undergrad curriculum is outdated and irrelevant for anyone who doesn't want to work in that particular space; while CS provides more *general* skills which are widely applicable. Best wishes for your apps

@ev a. If you don't mind, can you elaborate on why you want to sever ties with EE. Saying that CS is better suited is one thing, but saying you wish to sever ties with EE seems to suggest you had a bad experience. If you don't mind, can you share the reasons - I am curious. Also, are you by any chance a student at Caltech?

@ev a. You seem to have very peculiar reasons specific to you, not sure if anyone here can provide a good feedback. If you felt your calling for deep learning early enough, and lost interest in EE, you should have switched streams. At the very least, must have taken enough courses to cover for lack of exposure. If you feel anxious due to your lack of CS courses alone, it is understandable, but no point in fretting over it now. You'll get to know authoritatively in another 1-2 months. On the other hand, if you are anxious because ML/AI is very competitive at the moment - this is no secret. To be frank, I think you should have applied to another 3-4 programs. I regret not doing so myself. The deep learning professor who interviewed me said that applications to his lab nearly quadrupled from last year. Even if I do happen to get in, I may not get to work on the area of my choice due to fierce competition among admitted students. I guess everyone is anxious due to the situation being fuzzy and messy.

I assume you are interested in an MS and not PhD. If that is the case, why didn't you just apply for MS in EE? ML has very weak connection to CS compared to other areas. Much of ML is based on optimization and statistics which are integral to EE too. You could have just taken the appropriate ML electives and got away with an EE degree IMHO. This however changes if you want to do a PhD, in which case you should apply to the department of potential advisers.

Had an interaction/interview with PoI at UWashington yesterday (over phone). He said first batch of admits would be sent by this Friday (notified at the dept. level), and another batch would be done sometime around Feb 2nd week. Really nervous right now! He told me that he liked my application, and our interests match. But he didn't say outright "admit". Don't know how much to read into the conversation, but it is safe to assume that I will not get any sleep this week.

I've applied to their IDSS program, for transportation systems specification. I hope to work on either autonomous vehicles or intelligent transportation (smart cities project). I haven't heard back from them yet; have you?

@Yav Friendly Are you applying for PhD this year, or are you continuing at ETH? If you are applying, where to, if you don't mind me asking.

@Micecroscopy Thanks for the comments. I think it would be more appropriate to ask the inverse question - how relevant, conventional, or advantageous would a neuroscience based degree be for motor control, vision etc. in the context or robotics. How the CNS program is viewed in the engineering and CS communities. What drew me to the CNS program was that it isn't a conventional neuroscience program - majority of its faculty and affiliates are not biologists. I felt that it combined neural networks based methods for vision and control which is an exciting area. As you said, I am really hyped up for the interview, and would like to make it click. I'm still waiting for a few other programs though, which might be better fits - particularly ETH, CMU (RI), and UWashington. Best wishes for your applications!

Oh! This is funny. ETH is probably my top choice. I have applied to Prof. Raffaello D'Andrea and Prof. Andreas Krause there. Incidentally Prof. D'Andrea got his PhD from Caltech under Prof. Murray and Prof. Krause was a professor at Caltech till a few years back, before moving to ETH. I can't help but wish I was born 5 years sooner, that would have been a golden time!

@Yav Friendly Thanks for the info. Prof. Doyle is an absolute legend in control theory, and his robust control paper "State-space solutions to standard H 2 and H∞ control problems" (popularly known as DGKF paper) arguably marks the birth of modern control theory. But I don't think he is affiliated with the CNS program. Caltech also have another legend, Prof. Richard Murray who pioneered Networked Control Systems. However, both of them seem to work on systems and synthetic biology at the moment, as opposed to AI based robotic systems, which is a shame. For bio-enthused students, I think Caltech, UCSD, and ETH-Z are probably the top places. I will have an interview with Prof. Burdick who is their only robotics type control theorist (adviser of Jorge Cham, PhD comics!), and I plan on asking him about possible collaborations with other control theorists, machine learners, and vision experts. Keeping my fingers crossed Thanks again for your comments.

Haha, that seems decades away though. As an aside, I am absolutely convinced that control theory is one of the most awesome topics in the world. It has fertile connections to AI (reinforcement learning), synthetic biology (feedback circuits), and even economics! Systems and synthetic biology is certainly an exciting direction, with most control theorists (including my adviser) readying their guns to take aim. Though I am not directly interested in it (you can only choose one area, ), I am very excited for it, and hoping to keep myself up to date with the developments. In case it gets super hot, I'll probably jump ship Best wishes for your grad school applications!

Thank you @optogent and @Yav Friendly for your comments. I wouldn't say my field of interest is computational neuroscience, but it is certainly related and draws from it. I am more interested in the robotics end of things. How artificial neural networks can be used to make robots intelligent. For this, I am interested in mainly two parts: perception (mostly vision) and control. I think both have good connections with neuroscience. Till now, computer vision has not followed the path of biological vision, but my understanding is that the landscape is changing dramatically. Ideas from biological vision (particularly sampling) are taking over out in the wild type perception tasks useful for dynamic decision making. Control on the other hand is much more intimately tied to neuroscience for quite some time - particularly motor control and TD learning. However, I don't think I will be happy with studying "how does the brain perform a task". Rather, I would be happy to take inspiration from neuroscience, but I am more interested in "how to make a robot do a task", which may or may not imitate how the human brain works. I think I need to make it clear to the profesors. As you said, Caltech is a small place, and I can see myself working with only 2-3 professors. Though I am very happy with the interview invite, I am not 100% convinced if CNS is the ideal program for me. I really hope I click with the POIs though, since Caltech is Caltech!

Oh boy, the life sciences forum seems much more active than the engineering one. I just got an interview call (Skype, international) for Caltech CNS. My interests are more on the machine learning, computer vision, and robotic control side though. All my POIs have joint appointments with the CMS department. I just wanted to get a sense of where CNS stands in the neuroscience and natural sciences community. Caltech's CMS/CDS and EE programs are certainly top notch and probably top 3-5 in the world. Is the CNS program of similar stature, or are the MIT, UCB etc. programs well ahead?

Fall 2016 deadlines are long gone. I think you can apply only for Fall 2017. Post the question closer to September next year to get a better response from the community.

What is your gradschool plan? Are you planning grad school in statistics, applied mathematics, OR, CS? The answer will depend a lot on your interests, choice of grad school program, and career choice beyond. For example, I have heard quite a few say that for industrial big data roles, an ideal education background will be UG in Stats and MS/PhD in CS; or vice versa. If you wish to go a similar route, you might consider an MS/PhD in CS, for which a CS minor will undoubtedly help with admissions and doing well in grad school. On the other hand, if you want to do more mathematical statistics, then a Math minor might help more. However, I personally don't think this is a good option. As you rightly pointed out, you will end up taking many of the required minor classes as part of stats anyway! On your diploma and resume, it is better to put stats+CS (more variety) than stats+math (just more of the same). There are also other possibilities for gradschool like econometrics and biostatistics, for which a minor in economics or biology is a definite plus. In the absence of these, CS is likely to help more than math, since these are sufficiently applied disciplines which require CS skills. Full disclosure: I did a CS minor, so I am likely biased.

International student with background in *general engineering*. Interested in robotics, and related areas like control, machine learning, optimization, state estimation etc. No publications yet, but a couple in the pipeline. Hoping my advisers wrote strong recommendations. EE related places: CMU Robotics Institute Caltech Control & Dynamical Systems U Washington EE (Dieter Fox group) UPenn GRASP GATech ECE Best of luck to everyone

Firstly, what is your sub-field of interest? From what I have heard, evaluations are done very differently for different sub-fields. It is unlikely that you would be simply eliminated just because you don't have 1 or 2 required courses. However, you need to really show them something special to keep your app in consideration. For example, relevant publications, stellar recommendations, a very unique internship or work experience etc. Secondly, the type of courses that will be considered seriously depend on the sub-field. For example, if you are interested in robotics, having a few ME courses on control, kinematics, dynamics etc. could actually be more advantageous than having 4-5 courses on compilers, computer architecture, or operating systems. Having a low GPA is definitely not a positive whichever way you look at it. The extent to which it will hurt depends on the courses in which your grades were low, the university to which you are applying etc. Subject GRE may not actually add any value, but considering it is 80+% it will certainly not hurt. If you aren't concerned about a few extra $$, it's better to send. MS is almost always easier compared to PhD. But MS programs don't boost your PhD chances by a lot. Yes, your coursework problem can be taken care of, but the bar for admission will also be much higher for MS students. So unless you publish 1-2 top tier papers, a good PhD program will not be interested. Doing research during MS is not easy, and depends a lot on university. For small programs which take in just 20-30 MS candidates, chances of working with a faculty are bright. However, for large cash cow programs, most faculty aren't bothered about MS students. Funding is also essentially ruled out. So if you want to go the MS route, you must be very careful in your choice of programs. MS programs in Canada are research based, and cracking them is just as difficult as cracking PhD programs at a decent US school (top 30). From what I gather, CS PhD programs are currently extremely competitive. The amount of effort and aptitude needed to crack a 25-30 ranked CS program is comparable to cracking a top 10 program in most other engineering areas like ME. Added to this is the fact that you don't have the proper diploma. I'd suggest you have good backup plans like a CS MS program in India or be ready to shell some money for a CS MS at a top 30 US program.

Sci. Comp. programs don't really deal with optimization. Certainly not at the algorithmic or computational level. Some sci. comp. applications do use optimization, but it is mostly just formulating an very simple optimization problem and passing it on to a numerical optimizer (they don't write the optimizer themselves). My limited point is that there is very little overlap between the two programs. I have seen quite a few who applied to MS/PhD programs without exactly realizing the aims, topics, and scope of the programs, and became very frustrated later on. Just do your research well, and apply to the appropriate one. Best wishes

I just wanted to point out something. OR and scientific computing are very different fields, with very different course requirements, and very different career outcomes. Look at programs, their course listings, and most importantly alumni placement before deciding whether it is suited to your needs and tastes. OR is basically optimization and some applied statistics. The focus is clearly on problem solving and applications are primarily in business, economics, and possibly information technology. Sci. Comp. is totally different. There people view computation strictly as a tool to understand physical phenomena. You'll be taking courses on PDEs, finite element analysis etc. There could be some overlap in areas like stochastic simulations, but Sci. Comp. will use it strictly as a tool for something like computational material science. Think about what you want to do after MS - data crunching in banks, analytic firms, software, data sciences etc. are better served by OR. Anything involving CFD, climate modelling, computational biology etc. in places like national labs or R&D labs is Sci. Comp. domain.

From what I know, Operations Research and quite a few branches of Engineering (and Economics too?) will be happy to have a Stats MS in their ranks. The key is to communicate a convincing reason for the switch - considering that Stats is considered *hot* now.

I would vote for the statistics track, purely because it opens more options. Nearly all jobs open to OR majors would be open to Stats majors, but the converse is not true. However, if you have identified which type of job you want to get into (quant finance), then it makes sense to pursue the track best suited for the same which appears to be OR. However, do note that if you have a change of heart somewhere down the line, or if you don't like quant finance after a few years, Stats degree will open significantly more opportunities (e.g. in Healthcare) than OR. (PS: My brother was debating Stats vs OR, ended up applying to both, and enrolled in Stats).

Reopening this post. I have the exact same question with the exact same interests (planning, control, and learning). I'd really appreciate any advice or comments. Thanks