freshman in ML: how do you identify actually open research problems? [D]
Hi, I am a freshman who is trying to break into research.
I got into a well known university research lab in my country for the upcoming summer, and the prof said I am "better positioned than numerous others" for hardware-aligned machine learning topics. I am facing a couple of problems, and I would like to know how seasoned researchers deal with them:
How do you develop the intuition for what's open vs. what just looks open? When I look at a research space, everything either looks already solved or impossibly vague. There's no middle ground visible to me, yet. This bothers me.
How do you handle the feeling that every idea is either already done or not good enough, without it paralyzing you?
Ideas that I have "thought" of but have been done already: PQCache, async KVCache prefetching, roofline modeling for GQA decode phase.. etc.
A paper that says "future work includes X" BUT it is not the same as X being open, right? Someone may have done X last month and not published yet, or X may be open but intractable, or X may be open but require equipment which I don't have. I would have no way to know which. Morever the thing I want to work on might exist under three different names across three different communities, and if you search the wrong name you conclude it's open when it isn't. (LLMs with Web Search seems to help a bit)
Reddit threads that I have already looked into:
- https://www.reddit.com/r/MachineLearning/comments/1sayptq/d_physicistturnedmlengineer_looking_to_get_into/
- https://www.reddit.com/r/MachineLearning/comments/1nsvdqk/d_machine_learning_research_no_longer_feels/
- https://www.reddit.com/r/MachineLearning/comments/kw9xk7/d_has_anyone_else_lost_interest_in_ml_research/
My motivation to work on this field is to speed up ai-for-science initiatives, while making it more affordable.
[link] [comments]
Want to read more?
Check out the full article on the original site