My New Year’s Resolution – read a research paper every weekday

I'd state it with a bit more nuance.

From my experience in physics:

You can read a paper in 15-30 minutes and get the main points and contributions.

For good papers, you can spend multiple days to really understand the paper.

And for really good papers you can spends months to completely understand them, then extend them, and then publish your dissertation.

Reminds me of the 3 pass approach in the paper, "How to read a paper"

http://blizzard.cs.uwaterloo.ca/keshav/home/Papers/data/07/p...

I'm slowly finishing up my PhD and I tend to agree with you. However, he will probably be reading quite a different selection of papers than what you encounter during a PhD.

In my experience, time required varies quite a lot between papers and authors. If you take a paper from, say, top 1% (or 0.1%) of papers published on some broad topic, you can probably read through it once and get some relatively good understanding of what the author was trying to say (and very likely enough to write a short re-cap).

This is definitely not true for the remaining 99% of the published scientific literature. At one point I had the idea of keeping up with my field of research by setting up Google Scholar alerts. This resulted in about 1 paper per 2-3 days that seemed to be worth studying based on the title and abstract. Still at that rate I found it quite impossible to keep up with all the content that dredged up.

I have 2 comments saying just the opposite that good/bad papers require more time to read than bad/good papers.

So a bit of clarification to my comment. If I spend months reading papers about the same topic, the first papers will take me a lot of time but the last papers will take me less time because I have a lot of background and I don't have to check references, review concepts, understand maths proof or whatever. Probably those papers are very related and share a lot of information.

Good papers (for whatever you think it is good) require time to read because you really want to understand it completely. Bad papers (for whatever you think a bad paper is) require time to read because you cannot be sure if that paper is relevant for you unless you understand it completely. Although at the end you have a ton of papers to read and you check the abstract and decided whether spend time on the paper or not.

I've found that there's generally two levels of paper understanding. The first is achievable in a day, and consists of recognizing what the authors did and roughly how their technique works and why it's different. The second, much more nuanced level of understanding comes when you really dig into the paper and analyze every sentence, and commonly occurs when you need to extend or implement it. The writing is rarely wrong, but many things that are not relevant are either easier said than done or painful but boring implementation details. At that point, you could rewrite their paper from scratch if needed.

>resort simply to a high-level overviews of implementation and go straight to detailing their outcome

Those don't sound like research papers but articles... What Journal(s) are you finding these papers in?

I think it's almost impossible to regularly select only papers that will be helpful to your work or thought provoking in your field. The nature of the beast is that you're going to read a lot of irrelevant or simply bad papers so you have to develop strategies to discard those quickly. The other thing is unless you're following a specific thread of research from citations or a group of authors you're probably not going to know which papers are important/worth reading ahead of time.

Some strategies I've seen conference reviewers use and that I use to quickly evaluate whether a paper is worth a deeper dive:

1. Figures. Can you understand the figure from the caption or associated text blurb? Is it neat and labeled correctly? I've found that excellent figures are a strong signal that the authors put a lot of thought and work into the paper. If you look at Yann LeCun's early works they all have really excellent diagrams with clear descriptions. Some of his figures are so classic that they regularly appear in ML presentations to this day.

2. Read the related work section. A lot of people seem to skip this but it can be dynamite if you have knowledge of the field. You'll probably be aware of other papers referenced in this section and when the authors of the current paper point out differences and similarities to other methods this will give you a great idea of whether this paper's approach is worth understanding.

3. Save the Method and Result sections for last. For me, and I would assume most people, these heavily technical sections are very time consuming to understand and ought to be read very last. Read abstract, intro, related works, and conclusion first. I would bet I discard half of the papers I pick up without ever readin the middle chunk.

4. Conference or Journal it was published in. IEEE, NIPS, CVPR are all places where the best stuff in my area gets published. Papers not from top shops should have to do more to earn your attention. Strong figures being the one I usually go back to.

One thing I would not do, that I used to do, is worry about the number of citations a paper has. Awesome new research and old but potentially newly important work will not have many citations at first.

This is just my list. I'd appreciate other filters people have for reading papers!

A fun game is following the reference trail right back to the source works, often decades ago. Pick something popular, say convolutional neural nets. Do a brief literature review and try and read the most highly cited papers as far back as you can go. You'll usually hit some really interesting (and very readable) old papers that cover really groundbreaking things. Using the convnet example, at some point you'll come across the Nobel winning paper on how cat brains interpret visual stimuli (Hubel, 1963).

I used to have a subscription to nature and just leaf through it, going as deep as I got into each paper.

The subscription was quite cheap (may have been using student discount), and it has a sort-of "popular science" section in the front summarising the major papers for audiences from other fields. Nature publishes papers from all fields of science and it can be quite entertaining to take a deep dive into vulcanology. You'll also get to notice when different branches of science end up using the same methods.

Of course that's more for entertainment value (and possibly annoying your friends). for work I stsck everything that appears interesting into an app (app-tly named "Papers") and search through it when I think there could be something in there.

I like the idea of scientifically vetting foods, but as you said it might be hard coming up with a balanced viewpoint if you are not considering all the research. Also might be helpful to consider the impact factor of each paper you are citing.

A lot of CS papers are available on http://citeseerx.ist.psu.edu/

In CompSci it's pretty rare for there not to be open access versions of a major paper. Off the top of my head there have been a few in Nature but outside that almost everything ends up as open access.

You might be interested in Fermart's Library:http://fermatslibrary.com

Don't bother trawling through libraries. Look at the list of top conferences in your field and just go through the proceedings of those (e.g. for systemsy conferences look at the list given in the article as a good starting point). You should be able to get free versions of most papers you find through Google.

sci-hub?