I am about to choose ML research topic for my master thesis, but i am at a dead end. The problem is, that while reading research papers, i find solutions, but not an open problems.

For now, a came up with such ideas:

  1. Research neural network quantization algorithms - i started to research this topic 3 months ago and it seems to me, that current papers reveals high accucary methods. I have not idea what can i to do in this field
  2. Implement online video-processing system using quantized neural networks. It sounds also good, but it could be done with few lines of Tensorflow Lite code.
  3. Intelligent Image Database - image database, that has image classifiers and detectors for automatic annotation. It seems to be a good idea, but i need to find here a research part. For me this tasks sounds like pure engineering task
  4. Make some signal processing research. It could be connected with first idea as optimal hardware implementation. But i don't know where to search for suitfull specific task for ML, because a lot of dsp problems are solved with classical analytical mathematical methods.

As you can see, i always see some disadvatages in ideas. It stopping me from starting to work. Can you give me an advice how to develop one of this ideas to make it suitable both for engineering and research? Generally, how to search for research topics?


You're probably aiming too high: a research topic doesn't have to lead to a major breakthrough, and very often it's impossible to know what it leads to before doing it. A Master thesis is not very long so you need to find a research topic which is feasible within the time frame. A common mistake by Master students is to spend too much time on implementation, then botching the analysis and/or writing a poor quality dissertation because there's not enough time left.

  • Discuss with your advisor, they have the experience and they know what is a good topic.
  • Read a few random papers in your domain. Why? Because if you read only the top quality papers (yes, of course these are important to read), you can get the impression that you need to achieve the same level of quality. But that's not the case: these top quality papers are 1 in a million and are often done by people who have years of experience behind them. A "regular" research contribution is often just a small improvement or a decent analysis on a specific problem.
  • Typical process: just pick a general topic/problem you're interested in, then study the state of the art thoroughly (write your state of the art section at this stage), reproduce a few experiments and analyze the results: what can be observed? what are the limitations? what would be a reasonable idea to overcome one of these limitations? Then you implement and evaluate this idea.

Keep in mind that in research what matters is not the complexity of the topic or even the performance, it's the rationale and the method (and of course originality of the work, but even that doesn't matter that much at Master level).

  • $\begingroup$ You are absolutely right, i always aim too high , even if i don't want. Read a random paper is also a good advice, because i read papers written be expirienced researchers. Thank you $\endgroup$ – Дмитрий Feb 28 at 8:23

Not the answer you're looking for? Browse other questions tagged or ask your own question.