Dear Hackernews,<p>(respost because my other post didn't get a single response)<p>I want to do something with ML (DL in particular) and I'm in need of a dissertation idea (for my undergraduate course), ideally it'd be Computer Vision related / something to do with programming languages (with DL of course), does anyone have any ideas? It should include implementing a one (or more) papers and then extending them as well. I have a reasonable amount of compute available and the dissertation should take up a few months of my time.
There are still challenges in the field of real time interpretation of { American | Australian | etc. } Sign Language, eg:<p>(2022) <a href="https://medium.com/trends-in-data-science/interpreting-auslan-with-gesture-analysis-a-sign-of-tone-deaf-times-d26038017c1f" rel="nofollow noreferrer">https://medium.com/trends-in-data-science/interpreting-ausla...</a><p>(2005) <a href="https://link.springer.com/article/10.1007/s00138-005-0003-1" rel="nofollow noreferrer">https://link.springer.com/article/10.1007/s00138-005-0003-1</a> <a href="https://sci-hub.ru/10.1007/s00138-005-0003-1" rel="nofollow noreferrer">https://sci-hub.ru/10.1007/s00138-005-0003-1</a><p>You can overview the field, look at past techniques, aquire new training data from, say Deaf U and elsewhere, implement and extend.<p>Might take more than a few months ..
I suggest you look at papers with code for inspiration, just start exploring with different ideas or models and see where your experiments take you :) This is the best part of starting a research project, when you can decide your focus area or topic. I also suggest looking at posters around your department area (if they're pinned up) or at various conferences for ideas and inspiration.<p>What you're really trying to learn from a given example is not just the techniques used, but what exactly are the problems that need to be solved in a given research area, and what are the open questions in the field—this is where new research usually focuses. Secondly, you should aim to learn what sort of problem solving methods researchers use for your area of interest (at least that's how I would approach things if I were you).
Another possibility is ML for Cued Speech: <a href="https://arxiv.org/abs/2212.01083" rel="nofollow noreferrer">https://arxiv.org/abs/2212.01083</a>