4 ms·

Ask HN: What do you use to keep you update on ML/DL?

8 points by ReDeiPirati 7y ago

Hi everyone! What do you use to navigate-in-the-noise and keep you update in this field? Excluding HN which type of resources do you recommend to check regularly?

mlthoughts2018 7y ago
1. Read software libraries first. If new techniques make it into standard implementations of major libraries, it’s a solid signal the new technique is actually worth your time. 2. Don’t bother with arxiv / conferences / etc., until a paper is at least a year old and remains highly cited & investigated. In DL especially, papers are often slapdash permutations of existing ideas applied with one weird extra trick that ekes out “SOTA” results in one benchmark. Architectural choices are often not motivated by any evidence, and whatever scramble of grad student descent was needed to arrive at the final architecture choice is conveniently left out of the paper (and often would call the result into question from a multiplicity of testing POV). 3. Read what you enjoy. It shouldn’t feel hard to stay current in your main area. If it is feeling hard, you’re trying to force it. Maybe pick a sub-specialty you enjoy and only focus on that for a while. 4. Forgive yourself for not staying current. You are busy, reading research papers takes time & is not often the best use of time compared with competing interests. This is normal. If some situation / job interview / academic program is making you feel otherwise, it’s probably a red flag about that thing.
- croh 7y ago
  solid advice. thanks
codingslave 7y ago
I read https://old.reddit.com/r/MachineLearning/ https://old.reddit.com/r/MachineLearning/. It's not perfect, but most high profile papers will be posted there. The better ones will have a good number of comments, which helps me sift through the garbage. I browse it on a daily basis, never really reading the papers, just seeing what topics are being discussed most.
p1esk 7y ago
http://www.arxiv-sanity.com/ http://www.arxiv-sanity.com/ Usually just top hype/recent, plus specific keywords relevant to my research.

8 points by ReDeiPirati 7y ago

Hi everyone! What do you use to navigate-in-the-noise and keep you update in this field? Excluding HN which type of resources do you recommend to check regularly?

mlthoughts2018 7y ago

1. Read software libraries first. If new techniques make it into standard implementations of major libraries, it’s a solid signal the new technique is actually worth your time. 2. Don’t bother with arxiv / conferences / etc., until a paper is at least a year old and remains highly cited & investigated. In DL especially, papers are often slapdash permutations of existing ideas applied with one weird extra trick that ekes out “SOTA” results in one benchmark. Architectural choices are often not motivated by any evidence, and whatever scramble of grad student descent was needed to arrive at the final architecture choice is conveniently left out of the paper (and often would call the result into question from a multiplicity of testing POV). 3. Read what you enjoy. It shouldn’t feel hard to stay current in your main area. If it is feeling hard, you’re trying to force it. Maybe pick a sub-specialty you enjoy and only focus on that for a while. 4. Forgive yourself for not staying current. You are busy, reading research papers takes time & is not often the best use of time compared with competing interests. This is normal. If some situation / job interview / academic program is making you feel otherwise, it’s probably a red flag about that thing.

croh 7y ago
solid advice. thanks

codingslave 7y ago

I read https://old.reddit.com/r/MachineLearning/ https://old.reddit.com/r/MachineLearning/. It's not perfect, but most high profile papers will be posted there. The better ones will have a good number of comments, which helps me sift through the garbage. I browse it on a daily basis, never really reading the papers, just seeing what topics are being discussed most.

p1esk 7y ago

http://www.arxiv-sanity.com/ http://www.arxiv-sanity.com/ Usually just top hype/recent, plus specific keywords relevant to my research.