An Opinionated Guide to Machine Learning Research | Hacker News Reader

An Opinionated Guide to Machine Learning Research

joschu.net

165 pointshardmaru7 comments

This guy's time management is impressive. If you're diligently spending 1 day a week exploring new research methods, writing and reviewing your lab notebook, staying up-to-date on current SOTA research, writing SOTA algorithms for review, reviewing textbooks, engaging in "personal development" time (which apparently means general ML research) - when do you have time to actually do your main research?

One day a week hardly seems like enough to stay up to date on SOTA research (at least, anything more than being aware it exists).

There are over 200 articles on the arXiv in stat.ML from this week.

https://arxiv.org/list/stat.ML/recent

Not all 200 articles in stat.ML are SOTA research.

Certainly not, but an 8 hour day dedicated to deciding just that gives you only 2 minutes per article.

You only need a few seconds to read the title, and you can cut away 150 out of those 200 with that. If it's a paper simply not in your area of research, you don't need to bother. You can then read the abstract (30 seconds) for the rest, and finally you'll pick 5-10 papers to read. You'd skim them with 5-10 mins per paper, and if you like what you see in 1-2 papers, you'll give it a more through read. So this is 10 mins + 30 mins + 2 hours + buffer, which is at best 3-4 hours.

Of course this requires a strong filter at the first stage. You can also set aside time to investigate specific topics outside your work, but you won't do that as often as you keep up with your own literature.

re: goal-driven vs idea-driven

Feynman's algorithm

- keep a bunch of problems in your mind

- every time you hear a new solution, test it against your twelve problems to see whatever helps

and you can flip it around

- keep a bunch of solutions in your mind

- every time you hear a new problem, test it against your twelve solutions to see whatever helps

so to optimize this, what you should do is

- keep a half-dozen problems in your mind AND keep a half-dozen solutions in your mind

- then for each new problem, test it against your solutions

- then for each new solution, test it against your problems

https://www.youtube.com/watch?v=Z8KcCU-p8QA

Question to author: Why new and coming researcher has to be young? Why not an old aged person can do research. I see entry level research position blatantly discriminate based on age? If you have gender diversity then why not age diversity?

Reply on news.ycombinator.com