A rough guide on how to read academic papers in a structured way.
Reading academic papers is time consuming and sometimes frustrating. Information is dense, sometimes we lack context, or the paper is just not written very well. This guide is more of a checklist I developed through a bit of research online, on how to get through a paper efficiently, while extracting the most relevant content and understanding the concepts to the desired depth. Each step of the process has a bullet point list on what should be done for this step. If you have any further valuable tips on how to make the process even smoother, feel free to reach out to me and I will be sure to add them with the proper acknowledgements.
Before reading a paper in depth, we want to get a general idea of what the topic is, how it approaches the topic, and whether it is relevant for the research being conducted. The goal is to get the necessary context before making the decision to read the paper or not. To do so, we can read the title, abstract, and conclusion of the paper in question. While reading it can be useful to identify the keywords for future reference. After having read through those sections, it can be decided whether or not the paper is relevant.
Once we have decided whether or not this paper is of interest for us, we need to get familiar with more of the details. This can be thought of as acclimitization to the subject matter, which is the goal of this step. Knowledge acclimitization gives your brain the necessary context for understanding the rest of the paper in later passes. Once we know what to expect to a certain degree, our brain is primed to receive that knowledge and insights, making the absorption of the content much easier. To gain the necessary context before reading a paper, it is best to read the introduction and examine the tables and/or graphs.
The introduction gives an overview of the high-level goal of the paper and why it is relevant by:
The graphs and tables on the other hand, serve to get an understanding of what metrics are being used which helps to understand the contents of the paper better. They serve to provide support for the claims of the methods presented, thus give reader a better understanding as to whether these methods improve on current methods and to what extent. Furthermore, visual representation of data and performance enables an intuitive understanding of the contents.
After having primed our brains for the paper by giving it the necessary context, the next step is to do a full pass of the paper, which means reading it all, end to end. In this step, it is important not to get hung up on things that we did not understand immediately and to leave them for later. In this first full pass, the goal is to understand most of what the paper is presenting and finding out which concepts, terms, formulas, derivations, or algorithms need to be looked at into more detail. In order not to lose sight of the full picture, these areas that require more time to be understood should be noted down for reference in step four but should not be dwelled on for long in step three. To read the full paper, it is usually best to start by reading the abstract and conclusion again before proceeding with the rest of the text and to take some quick notes per sections with short breaks in between the sections.
In the final pass, the goal is to look into all unfamiliar terms, algorithms, definitions, concepts, and methods that were noted during the first full pass. It’s important to take enough time to thoroughly understand everything in this step, so that we can conclude the reading of this paper with a confident feeling of having understood the subject matter. Use any external resources that can aid in understanding such as: presentations, blogposts, articles, textbooks, etc. Also, refer to the cited papers for help finding the appropriate resources. Finding the right aids is the key factor in successfully completing step four.
Some useful places to look when it comes to machine learning research:
The Machine Learning Subreddit The Deep Learning Subreddit PapersWithCode Research Gate Machine Learning Apples
Top conferences: NIPS, ICML, ICLR, CVPR
Especially when delving into a new topic, in can be very helpful to solidify your knowledge by writing short summaries on the topics or papers being read. This practice really cements the new topics by employing the feynman technique, where you attempt to draw on your knowledge to explain a novel concept in a comprehensive way to someone else. Writing summaries has the added benefit that they can be shared online through blogposts and serve as a reference for others when reading through similar material.
Reading research papers can be cumbersome, especially when you are new to it or to the topic. Having a structured approach can make it more manageable to take in the contents of more challenging papers but it is also important to note that knowledge acquisition takes time and should not be rushed. So take your time, remain focus and be sure to revisit more difficult concepts multiple times.
To summarize, these are the steps recommended when reading a research paper:
https://developer.nvidia.com/blog/how-to-read-research-papers-a-pragmatic-approach-for-ml-practitioners/
https://www.turing.com/kb/how-to-write-research-paper-in-machine-learning-area