of Computer Science
FAQ - Transferring into the CST from another Tripos
-
Question: Can I change Tripos to CST?
Answer: Unlikely.
Cambridge has the entirely admirable aim of making transfer between Triposes possible. While this is a great ideal, and can work very well depending on the source and destination, transferring into CST from essentially anywhere is fraught with difficulty, and we strongly advise you against it. The latter point is not widely appreciated and the following notes aim to clarify matters.
For completeness, the Ordinances governing such transfers can be found here:
https://www.admin.cam.ac.uk/univ/so/pdfs/2021/ordinance04.pdf
-
Question: So, what's the problem?
Answer: There are several.
-
Once upon a time --- well, around the 1950s --- it might have been reasonable to consider CST as a backwater of mathematics or physics. In this historical hinterland it was probably reasonable to expect an easy transition. Unfortunately, while still prevalent, this attitude has been unrealistic for several decades: CST is now an entirely separate subject.
-
There is also a tendency for people to assume that "having done a bit of Python" makes them a Computer Scientist. At our current moment in history, "having done something cool using Tensorflow" seems to lead to similar claims. Sorry, but no.
-
Let's say you're currently following the X Tripos. Transferring to CST Part 1A (without starting from scratch, a possibility which is discussed below) requires you to catch up with some proportion of a year's worth of programming (two courses with very different languages), electronics, algorithms, databases, graphics, operating systems and so on---the details are all here:
https://www.cl.cam.ac.uk/teaching/2122/part1a.html
Do you genuinely consider this a possibility? Also...
-
If you intend to transfer into CST Part 1B you have an entire year's worth of material to teach yourself. Ask yourself: do you think you could have self-taught the entire first year of your present Tripos over a single summer, without any support in the form of lectures, supervisions and so on? (And no, we would not be providing extra support of that kind.)
-
If you're planning on entering CST Part III you have a single summer to catch up with three years of material. Could you teach yourself three years of your current Tripos in a single summer, with the same constraints as mentioned in 4? If the answer is "no" you should probably reconsider.
-
If you really want enforced honesty about your chances, try some exam questions from here:
https://www.cl.cam.ac.uk/teaching/exams/pastpapers/
and see how you perform.
-
It used to be the case that transfer from NST Part 1A into CST Part 1A was more feasible. This was because CST was one of the optional bench subjects in NST Part 1A, so students taking that option would be at a much smaller disadvantage when transferring. That option is no longer available.
-
In recent years the Department of Computer Science has experienced a huge increase in applications to the CST, to the extent that physical space has become a limiting factor. As a result, the subject is "numbers limited" in terms of applications, but has in some recent years also applied an outright ban on certain transfers. We have no control over the latter.
-
-
Question: OK, in any case, what are the specifics for transferring into CST Part 1A?
Answer: If you are already following another Tripos we are very unlikely to allow you to transfer within the same year, for the reasons discussed in 3 and 4.
We might however invite you to be tested (TMUA and CSAT tests) and interviewed alongside next year's applicants. Why is this?
-
You already have a place at Cambridge. CST is hugely over-subscribed. We may not in fact be able to accommodate you for the reasons given in 8.
-
You already have a place at Cambridge. If we allow you a place in next year's intake, then you will quite possibly be excluding someone who does not currently have a place at Cambridge. Some might consider this at least impolite. In order to justify it we need to be sure that you are (a) someone we would make an offer to in any case and (b) sufficiently good to justify us excluding a normal applicant.
It might be worth noting that in recent years two people have opted to try this, Neither would have been admitted had they actually applied in the same year, and neither were allowed to transfer.
-
-
Question: What about transferring into CST Part IB?
Answer: See 4 above. Your chances of convincing us that you will be prepared are extremely small. So while the regulations allow it in theory, we will almost certainly consider you unprepared unless you can offer actual evidence that you are. (Assurances of hard work to catch up do not constitute actual evidence.)
-
Question: And CST Part II?
Answer: In my reading of the Ordinances, which for Part II differs from Part IB in that the phrase "...or in another Honours Examination..." is missing, you can't:
https://www.admin.cam.ac.uk/univ/so/pdfs/2021/ordinance04.pdf
This seems pretty explicit: you only get to do CST Part II if you've already done CST Part IB. You might feel like arguing that the Ordinance isn't clear about whether the intended interpretation has an implication or a bi-implication; however in any case, refer to the previous answer: you would need actual evidence that you've somehow self-taught two years of material to an acceptable standard.
-
Question: I think I'm starting to get the message. How about CST Part III?
Answer: This option turns out to be a bizarre edge case.
You can in theory (and there is at least one precedent) transfer into CST Part III. This is despite the fact that it is an advanced course continuing from Part II, and not a conversion programme.
We do not recommend this. At all. Not even a little bit. You are strongly advised to find a Computer Science conversion Masters at another University.
The details are here:
https://www.cst.cam.ac.uk/teaching/part-ii/part3info
and you will see that it is your Tutor that you must contact, not the Director(s) of Studies for CST. (Although we will be asked what we think.)
Why is this an edge case? Well, on the one hand it assumes that you have covered the material up to Part II, and are good at it, which is why even those who have done so need to achieve a First. However, Part III allows a lot of specialization. That means that if you've attained sufficient expertise in a narrow area of Computer Science then it may be possible to choose a sufficiently narrow collection of courses that you can succeed.
You might want to consider though what happens if you do this and Graduate with a Masters in Computer Science, while lacking most of the fundamental knowledge that a Computer Scientist should have. (A future employer might rightly be upset when it turns out that you're great at Category Theory, while committing regular crimes against good software engineering.)
So: if you're sufficiently well-schooled in a narrow area and can convince us that you can succeed at a correspondingly narrow set of courses, and also that you can self-teach enough of Parts IA, IB and II to have a fighting chance, then we may sign off a transfer application, but there is one final point to note: the College DOSs have very little involvement in Part III, which is almost entirely run by the Department. As a result, it is entirely your responsibility to be prepared. We will not be able to offer remedial support of any kind, and we will not be arranging extra supervisions to help make up any deficit.
We really do advise strongly against it.
-
Question: Anything else I should consider?
Answer: Yes, are you sure you feel that this is an appropriate way to get a place on the CST?
As already noted, you already have a place at Cambridge. By accepting a place on your current course it is entirely possible that somebody lower down the admissions ranking got pooled, simply due to space constraints. Having been pooled, it is also possible that they ultimately were denied a place at Cambridge. Had you applied for the CST in the first place they might well have received an offer. This is not an admirable way to behave.
FAQ - General questions about AI
"I confidently expect that within 10 or 15 years we will find emerging from the laboratories something not too far from the robot of science fiction fame.” --- Claude Shannon, speaking in 1961.
-
Question: When will General AI be achieved? That is, when will AI operate at a genuinely human level?
Answer: Some time in the future.
-
Question: Eh? Seriously? What sort of an answer is that?
Answer: In short: look at the quotation above. If a giant like Shannon can get the answer wrong then so can I. I stopped giving a more explicit answer when I realized that it just got me trolled.
Let me elaborate. The existence of intelligent systems is not controversial: by definition, we are the proof. There is of course the possibility that we are too complicated to understand and reproduce, in which case General AI is out of reach, but I find that possibility very unlikely. So: "some time in the future" is a safe bet.
It's coming up with a timescale that's tricky. First, brains are ridiculously complex - far more so than anyone thinks until they learn some neurophysiology. Second, nobody has any idea how all the really interesting stuff - the stuff like consciousness, that philosophers get excited about - works. At this stage of the game you can take any Great AI Achievement that you like, but in the context of what brains do it's incredibly limited, despite being a massive achievement in AI terms. So: I'm not about to guess at a timescale, although in the past I've claimed "not any time soon" and that's as far as I'm going!
The FAQs on this page are an ongoing effort, come back later for more...
FAQ - PhD applications and Internships
-
Question: I am an undergraduate and I want to do something during the summer. Can I come and work with you for a few weeks."
Answer: If you are currently a student in Cambridge and there is a grant available to pay you, then perhaps this will be possible. I do not currently have grant funding for this myself, but it may be the case that your College or someone else has a suitable scheme; if that is the case then feel free to contact me.
If you are not currently a Cambridge student then I am unable to support this kind of internship. Please do not send a long email explaining how fascinating my work is, with an attached CV, that you've also copied to 200 other recipients.
-
Question: Are you accepting PhD students.
Answer: Yes.
-
Question: Er, OK. What else should I know?
Answer: I aim to take on a couple of people per year.
-
Question: Are there specific areas you are interested in supervising?
Answer: My main research interest at present is in the application of machine learning to automated theorem proving. So if you have a background in either and find the idea interesting that's great.
(If you're good at both that's even better!)
-
Question: How about machine learning more generally?
Answer: Yes, I will also consider this. I am generally interested in machine learning, with a strong bias towards principled approaches founded in mathematics/probability. I have in the past in particular worked in computational learning theory and Bayesian inference and am still potentially interested in supervising such things.
-
Question: I want to apply #thing-that-is-deep to #some-problem. Will you supervise me?
Answer: Not without a lot more information. You are describing an approach that can lead to a PhD, but probably not a hugely interesting one. (Do you really want to spend 3 to 4 years writing Python scripts to call Tensorflow?)
-
Question: BUT! BUT! DEEP STUFF IS REALLY COOL!
Answer: Indeed it is. But unless you're proposing something beyond straightforward variations on known architectures and/or straightforward application of such architectures to some specific problem, then it's not a recipe for an interesting PhD.
-
Question: So why is everyone doing Deep Stuff?
Answer: Because people love bandwagons. Why? Because bandwagons attract funding, whether that's a good idea or not. You should be very wary of jumping on bandwagons, unless you got there first. This is because any area with a bandwagon will have a small number of early joiners leading the field and doing the interesting stuff, and a very large number of people doing derivative things of limited importance. The real problem for most who join is then that there is so much stuff being published that it's extremely difficult to make a mark.
There are of course plenty of things in the Deep Stuff world that are worth pursuing and potentially extremely interesting. The problem for a potential PhD researcher is that these also tend to be hard, and a responsible supervisor might well take the view that you should not work on them because of the risk that you will make insufficient progress to obtain a PhD.
(And, yes, the supervisor might be wrong, and you might do something amazing - such things do happen. This is not an exact or predictable process, but it's worth being aware of these many considerations.)
-
Question: Are bandwagons really that bad?
Answer: Yes! And anyone who has been a researcher in AI/ML for long enough will know how the dynamic unfolds.
For example, consider a bandwagon from a decade or so back: Evolutionary Computing. What proportion of the huge volume of published research in this area has made a genuine impact? (Answer: there is a small core of really useful material, but that's it.) How much of it do you see in NeurIPS/ICML/AAAI these days?
And this is not claimed with the benefit of hindsight: plenty of people predicted it.
-
Question: I am interested in #ai-related-thing and would like you to supervise my PhD. My CV is attached. Please tell me if I've been accepted.
Answer: If I do a web search for "cambridge computer science admissions" I immediately find:
https://www.cst.cam.ac.uk/admissions
This comprehensive resource explains how one applies for a PhD place. By not bothering to read it you mark yourself as someone who really isn't serious about applying.
If, as is often the case with such emails, #ai-related-thing clearly has only the most tenuous relationship to my stated interests, then the preceding observation is even more pertinent.
If, as is also quite common, you have obviously copied the same message to 50 other AI/ML researchers, then you really do need to question your approach...