4 ms·
There is an alternative approach that inadvertently worked for me that may work well for you, particularly if your goal is to learn things beyond a conventional
by jandrewrogers 13y ago
There is an alternative approach that inadvertently worked for me that may work well for you, particularly if your goal is to learn things beyond a conventional computer science curriculum. I have been working in high-end theoretical computer science for many years, with significant advances to my credit, but I went to school for chemical engineering. People are often curious as to how I acquired so much esoteric domain expertise in theoretical computer science without spending any time in a computer science class.
Pick an interesting "hard" problem in computer science and attempt to solve it. This works best if either there are no published solutions (e.g. massively parallelizing ad hoc graph search) or the solution space varies widely under the possible parameters (e.g. the design of a unique high-performance database kernel). Iterate, be creative, search the literature for interesting ideas that you can borrow from, and evaluate new designs you come up with. If something does not work, understand how and why and try to fix it. Wash, rinse, repeat.
This sounds like a lot of work, and it is, but there are three unique upsides. First, the challenge of it can be quite a bit of fun in its own right and there is a clear goal of what you are trying to achieve which helps keep you motivated and measures progress. Second, you will cover more of the phase space and explore some tangential theoretical areas while hunting for solutions to narrow sub-problems that you would never be exposed to in a more structured setting. Third, you will invariably explore some unusual or unorthodox ideas that you simply would not be exposed to in a directed, formal program. From the standpoint of being an effective theoretical computer scientist, it will help you develop a unique perspective of the problem space that is different than the perspective that arises from more structured programs.
Also, you might actually solve one of the unsolved theoretical computer science problems you have as your challenge in addition to developing a deep understanding of the surrounding problem space. That happened to me (and then replicated later once I realized it was possible) and it would never happen if I was simply taking courses on the subject matter.
This was an effective way for me to gain in-depth understanding of a diverse range of theoretical computer science areas on an informal basis. One of the downsides is that your rote knowledge will not perfectly overlap what is taught in CS curricula even though what you do know will likely be more valuable (e.g. in spatial indexing some CS courses teach priority trees even though they have no memorable theoretical value).
- eli_gottlieb 13y agoPick an interesting "hard" problem in computer science and attempt to solve it. This works best if either there are no published solutions (e.g. massively parallelizing ad hoc graph search) or the solution space varies widely under the possible parameters (e.g. the design of a unique high-performance database kernel). Iterate, be creative, search the literature for interesting ideas that you can borrow from, and evaluate new designs you come up with. If something does not work, understand how and why and try to fix it. Wash, rinse, repeat. How do you verify your attempted solutions and talk to other researchers? I've always found it hard enough getting a hold of my own advisor or someone else in our department, let alone someone at an official research institution if I weren't in one.
- jandrewrogers 13y agoThat part has always proven to be pretty easy. For most interesting problem areas of theoretical computer science, there is rarely more than a handful of deep experts. Anyone else you talk to would likely need to invest a lot of effort understanding the nuances of the theoretical problem before they could have an informed opinion, which they aren't going to do. Identify the few real experts on the topic and talk to them. These individuals have two useful characteristics. First, they can identify 99% of non-solutions immediately upon casual inspection. Second, if you show them something novel in their area of expertise that they cannot immediately dismiss, it piques their interest and they will invest the time to actually understand what you've done. I've never run into a case where I developed a novel algorithm or similar where the other experts working in the field were not happy to study it if I sent them a cold email. However, a big part of the learning experience is being able to show that a particular algorithm must have a desired set of characteristics solely by its description before you show it to another expert. The unique skill you are learning by doing things this way is how to take an arbitrary set of complex behavioral constraints and designing an algorithm or system that conforms to those constraints. This is very different than the more orthodox course approach where you learn the constraints within which a given set of well-known algorithms exist and then select the appropriate algorithm for a use case. There are many real-world cases where all well-known algorithms cannot fit within the necessary constraints. Designing custom algorithms is not particularly difficult, but it requires grokking all the theoretical knobs that can be turned for a problem space rather than memorizing variants of a particular solutions.