When I read research papers about machine learning, clustering ... I wast a lot of time in reading one paper, and usually I find that finally it is not useful or convenient for what I wanted (even if the abstract looked interesting). How do you usually organize yourself when reading papers and taking notes of them ?

asked Feb 23 '12 at 04:36

shn's gravatar image

shn
462414759


5 Answers:

You really need to learn to read a paper on many levels, and most papers you read will be only in a very superficial level. Force yourself to skim a paper in less than 5min, and then make yourself answer, in writing, what are things like the motivation, key ideas, main related work, main result, etc. Deeper levels of reading include being able to implement a silly version of the paper's algorithm (if there is one) or convince yourself that intuitive special cases of the theorems they proved are true. Learn to not skip equations, but also know that most people don't read every symbol in the equation but pattern-match it with other known similar equations (same thing with algorithms). Learn how to do that.

answered Feb 23 '12 at 20:02

Alexandre%20Passos's gravatar image

Alexandre Passos ♦
2554154278421

Yes, I think that to skim a paper in about 5min then read it in a superficial level. If it is as interesting as what I expected I read it with a deeper level.

(Feb 24 '12 at 08:31) shn
Read the Abstract 
read the Results & conclusions 
if ( interesting and relevant)  {
        copy into Mendeley (http://www.mendeley.com/)
        read the rest of the paper.
}

answered Feb 23 '12 at 07:28

Dov's gravatar image

Dov
942410

Is mendeley really useful ?

(Feb 24 '12 at 08:28) shn

I use BibDesk. Its database is a BibTeX file, so integrating with LaTeX is a no-brainer.

(Feb 24 '12 at 11:16) Kevin Canini

Even I am facing the same problem , but what I usually do is , read first the asbstract ,their main contributions and some of the mathematical equations and the conclusions. If this answers some of my questions, I usually read the whole paper.This may save your little bit of time.

Here is a link for how to read and organize scientific paper.

answered Feb 23 '12 at 05:02

Kuri_kuri's gravatar image

Kuri_kuri
293273040

edited Feb 23 '12 at 05:08

I hate to add an answer to an otherwise well answered question, but one thing to remember about reading research is that there is a large number of crap or uninteresting papers being published every day.

The reason for this is simple -- publish or perish -- but as researchers, everyone suffers because the signal can get lost in the noise.

This means you must be vigilant when reading research. Can't work out what's going on after reading the abstract? Don't bother. The results are at best incremental using a weird evaluation metric? Don't bother. Very poor communication skills? Don't bother. Paper from 2000 in an active field, but has no citations? Read a little, but don't stress too much. It may sound a little pretentious, but the reality is that we have too much to do to waste time reading a paper from a conference that the authors just put in to bump up their quota.

It's also important to keep track of all the papers you ignore, and why you ignored them. While Dov only puts relevant papers in Mendeley, I put everything in there and try to keep notes on most of the papers (I've been a little slack, but it's a low priority task at the moment). I do this because I kept finding myself getting halfway through a paper and realizing I've read it before.

answered Feb 23 '12 at 16:26

Robert%20Layton's gravatar image

Robert Layton
1625122637

In this context, what do you think about this project ?
http://www.mydivvi.com/

(Feb 23 '12 at 17:07) Dov

also interesting: http://techcrunch.com/2012/02/05/the-future-of-peer-review/?utm_source=feedburner&utm_medium=feed&utm_campaign=Feed%3A+Techcrunch+%28TechCrunch%29

(Feb 23 '12 at 17:07) Dov

Here is my personal workflow to complement Dov's:

read the Abstract 
read the Results & conclusions
find the plots with the test set curves / benchmarks
try to dig in the text for info about the number of samples, features, runtime and memory usage
if (interesting and relevant and scalable)  {
   save a copy in dropbox
   if (2 columns layout pdf) {
       paper2ebook the pdf
   }
   send pdf to my kindle for reading when bored in public transportations
   mentally add to my list of stuffs to implement in a WIP pull-request for sklearn
   (or more frequently try to convince someone else to do so)
   (and often, completely forget about it because of paper overload...)
}

paper2ebook is a tool to make 2 columns papers somewhat readable on mobile devices.

answered Feb 25 '12 at 12:37

ogrisel's gravatar image

ogrisel
498995591

edited Feb 25 '12 at 12:41

+1. thanks of the paper2ebook.

(Feb 25 '12 at 14:19) Dov
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.