Sunday, September 4, 2022

// // Leave a Comment

PhD system and duration: 5 years with 3 years planning - isn't it too long and risking obsolescence on the go? What about contributing to many projects and not just one "personal" thesis?

This is a continuation of the 2008 article:

A Start-up or a PhD? - that is the question

The question-essay was asked to the comments section of a Q&A session video for clarification of the PhD program of  (...) [but deleted by the channel owner soon after]

Is it possible to plan your research for 3 years in a PhD program and what if your plan gets obsolete which is highly likely in the current speed of innovation and competition? Aren't 5 years too long and isn't working "on one problem" for 3 or 5 years, the PhD standard, an artificial limitation for students' development?

 Three years or 5 years with the exploration phase may be common term for a PhD, and working on "one-problem" (whatever "one" means as the topics, problems and fields overlap), however I wonder isn't that too long given the lightning speed of innovation and the speed of "repainting" the AI landscape, especially regarding the "planning" aspect, which is one of the requirements (for any PhD program)? On a broader ground, I remember an interview with the physicist Freeman Dyson (indeed he mentions, that he lacked a Phd...), on Youtube: "Why I don't like the PhD system", 1:38.

He "didn't like" the 3-year PhDs term in Cornell, because for him it was working for too long on one project (with the students) and it was too limiting for the students as well. He said he rather preferred one-year term, thus working on three projects for these 3 years. Prof. Vechev mentioned, that the nature of the institute encourages collaborating on others' project and it provides a rich environment for self-arranging lots of seminars and meetings between all the researchers, which allows enormous interdisciplinary learning rate. However, it doesn't solve the following automatically:

What if the plan of the student and his supervisor gets obsolete on the go? I guess it's a well known phenomenon that if an idea in CS and application programming (or maybe any domain) is not developed to some form of "completion" - developed, published, implemented - in a timely manner (with unknown term, depending on the other researchers and companies around the world), somebody else will do.

Others probably had the same ideas or plans even before, or they would have them soon. Similar with anything in R&D, as one aspect of general intelligence is its convergence: it's a systematic exploration of the affordances: if something is thinkable and doable, sooner or later it will be done. There is no central organ to distribute "the ownership" of the ideas and projects, so I guess you have experienced such overlaps and competition in your research. We see similar "state-of-the-art" being produced from different sources in about the same time, as all have similar background and goals and interact with each other.

How do you solve that, is it/has it been a problem? You talked about the normal topic-shift while initially searching for the best direction for a student, however that's the beginning; then there is a 3 year period with planned work, where a sudden topic-problem-goal-"obsolence" could crash any rigid structure and expectations. 

E.g. the plan of the student Ivan is to solve, say the automatic solution of programming contest problems*, which prof. Vechev mentioned in a podcast.

He will apply some Neural-Turing machines, combined with techniques of DeepCode etc. However DeepMind or some new "Perelman" happened to solve it 1 or 2 years before the end of that dead-line with a similar method, or with another or a more advanced one etc. and with a higher precision.

One reasonable path is probably to extend/rework the project to build upon the other project(s) etc., if it's doable. 

Or I assume that these "plans" are actually flexible, because I don't think there could be a reasonable many-years-long plan for true R&D, with fast learning rate and where real breakthroughs could happen, either from the researcher-himself or from the whole world.

If you can plan the content and the results for 3 years ahead with a high level of confidence and detail, I think it sounds like you already knew the results, i.e. it's less of an exploration "in the unknown" and more an implementation, i.e. engineering; and even in the latter more predictable domain, in CS it's usually hard to make precise predictions for the required time to develop the solution, especially when implementing something for the first time, there's a lot of both "known unknowns" and "unknown unknowns". Etc.

Thanks for the QA session and good luck!

* Note, 10.4.2024: That really happened to some extent: they did solve it, soon after the comment, with AlphaCode, LOL. 
Also this comment, turned then into article, was removed by the INSAIT channel's maintainers.

0 коментара: