Which papers to read?

Last update: Mar 10, 2015

Note: I am junior researcher. Ideas of this page are shallow, will improve by time.

Papers Source?

  • Sure only from top tier conferences and Journals. See.
  • Any other source will have fatal problems...it will waste your time.

Why do researchers cite papers?

  • It worth first to understand the reasons behind citations. In order of the frequency:
  • Literature Review Citing
    • You are solving the same problem I am tackling. [I may cite you]
      • Your solution was an important literature jump. [I probably will cite you]
    • You are solving the same problem I am tackling in a very related way. [I probably will cite you]
    • Doing survey and need both above types.
  • I Use you
    • I need algorithm of problem and you solved and provided it.
    • You have nice dataset
    • You provide some nice tool
  • Your Good Results: I need to compare to you or refer to your old nice performance
  • Inspiration: Your work inspired me to come up with an idea.
  • Building block for most of people: Many people depend on it some how. E.g. SIFT work in vision.
  • and some other reasons :)

Dependency Graph?

  • Imaging building graph, where nodes are papers
  • Let edge A-B: Reading B really somehow helped A to formulate its solution.
  • I think this graph will be sparse.
    • That is many papers are nice one, but if we removed them from literature, we are all fine.
      • E.g. Christophe 2008 ESS Algorithm is nice one. However, it is only cited as it is nice one. No one will depend on it in his work.
    • Some work is not important in anyway.
    • Some work is not practical. E.g. you need 10 minute to classify an image.
    • And so on...many reasons why many papers are not that important. But we all like our work to be published :)

Depth and Breadth

  • One way to read a paper is through multiple iterations.
  • In each iteration you gain more and more information.
    • May be first iteration is 10 minutes. Realizing the problem, contribution, some results
    • Second iteration go more in contribution details, little more in results
    • Third iteration may be understand the heavy details
    • May be now criticizing.
    • And so on
    • The more you go, the more depth you do.
  • There are thousands of papers. We can't read all of them. We can't understand every one in depth.
  • Normally, we have breadth in papers (1-2 simple iterations on many papers)
  • We have depth in what we care with it more.
  • The art is for each paper...to which depth to go?!

What do I need from literature?

  • Understand the different directions to solve a problem.
    • We don't need to know paper in depth. We need breadth of papers.
    • Just realize the different approaches
    • We need to iterate on papers and cluster them.
    • We need to mark the promising ones to go little depth
  • Realize what are the state-of-the art work in the previous little years
    • This is critical. You need to understand that and reasons behind it. These papers are read in depth.
  • Building block papers
    • Suppose the previous 2 steps will clearly point out to these papers.
    • They may be recent. They may be old (E.g. Lecun 98 in CNN).
  • Inspiration
    • Many times, nothing inspire us. As it needs in-depth reading, thinking and criticizing.
    • We usually depend on our imagination to find new solution.

Strategy

  • Don't over estimate the importance of reading literature papers
  • Survey may tell you about the different directions to tackle a research problem.
    • List the papers they refereed in each direction.
    • Pick 1-2 papers with the highest citations and read them in depth.
    • Later, iterate in a breadth style on other papers in each direction and see which are more promising. Add some depth for these ones.
    • When you decide which direction to take in your research. Go more depth to literature in this direction.
    • Through your research trip, keep reading in different directions...increase depth.
  • Identify state-of-the-art work in recent years.
    • In vision, we have some competitions. This is one way to know who is the best cross years.
    • Generally, papers will refer to them.
    • Read them in depth. You are competing with them.
  • New conferences and Journals
    • In each year, do your iteration and decide depth/breadth for the new directions / approaches.
    • Within range of each ~5 years, new things are in the spot (e.g. Deep Learning in 2012). Who realize that early and orient his research toward it will lead in the research or be remarkable.
  • Highly cited papers
    • There will be some papers in each field with a great citation record. Everyone refer to them to some reason. It worth having some depth in them.