“Reading research papers is something most PhD students learn via osmosis”. This article will summarise an approach by Prof. Andrew Ng on reading academic literature.


Approach

Consider a project-based scenario where in order to accomplish a goal, there needs to be some research performed on a subject before a possible solution is designed. The objective is to gain insight on past proposals, latest developments and existing challenges in the subject.

Perform the following steps:

  1. Compile a list of papers (arxiv, Medium/blogs, GitHub)

  2. Read around the list. The progress of the articles often looks like this:

    Research Progress

  3. The following can be used as general guidelines to the extent of research and what it would imply

    1. 15-20 papers: This implies a basic understanding of an area (eg. speech recognition).
    2. 50-100 papers: This implies a very good understanding of an area (cutting-edge or close to mastering).

Reading a Single Paper

A single paper is read in multiple passes, this is done as all papers are split into sections which could be useful for understanding the objective and results of the research.

A bad way to read a paper is to start from the first word and end at the last. - Andrew Ng

  1. First Pass: Title, Abstract, Figures

    This is enough to get a good sense of the problem statement and a general overview of what the paper is about with minimal reading.

  2. Second Pass: Introduction, Conclusion, Figures, Skim the rest

    Usually the abstract, introduction and conclusion gives a very clear summary of the purpose of the paper. After this, skim the rest of the paper. Generally the related work section is difficult to understand and usually filled with a lot of different work that may not entirely be relevant.

  3. Third Pass: Read everything but skip the math

    It’s easier to gain a theoretical understanding before deep diving into the math that makes it work (think Feynman technique). Go for the efficient high information content first and hard material later.

  4. Fourth Pass: Read everything but skip parts that don’t make sense

    Sometimes scholarly articles are published at the boundaries of our knowledge and bleeding-edge research at the time might be irrelevant in the current day. These parts can be skimmed or more time could be spent if mastering/deep researching is the goal.

Summarise