The world’s leading publication for data science, AI, and ML professionals.

How You Should Read a Machine Learning Paper

Keys to Read and Comprehend a Paper The Right Way

Introduction

Throughout this post, we will review the most important principles you should take into account when reading a machine learning paper and if you actually need to read papers to advance in your path as Machine Learning Engineer/Practitioner.

Picture from Unsplash
Picture from Unsplash

What is a Paper?

A paper is a written article, but not a journalist one or one you could find in an internet blog but an academic article. It is a document that externalizes the results of an academic investigation and that presents to the reader all the necessary information to argue and reveal the discoveries found by the researchers.

With its final aim of increasing the knowledge field, the scientific paper acts as the written medium by which the knowledge is officially expressed, spread, and reviewed by the community. This last point is crucial, a scientific paper to be considered as such has been reviewed and verified to assess its quality.

A paper that has not been reviewed and published is called preprint and although it may be available to read, you should take special care and read it with an even more critical eye.

Once a paper is ready to be published, it will be done through an editorial that will take care of adding it to a scientific magazine. This is a common metric to assess the academic relevance of a specific researcher: How many publications does he/she has in prestigious magazines, as well as how many citations has his/her papers have had. Some of the most prestigious magazines are Nature, Cell, Communications ACM, Science…

To be able to access this scientific knowledge, sometimes you have to pay for it, and not a small amount. Normally, the most common clients of scientific magazines are researchers and universities. If you are currently studying at the university, you can ask the library how to access these contents for free.

If you cannot access scientific papers freely, you can always use Sci-Hub to do so :).

Lately, the Open Access movement has been pretty popular, its goal is that everyone can access papers for free and there are paper repositories like arXiv which has pdfs of thousands of papers with different versions of them. Be aware though, here you can only find preprints.

The objective of these platforms is to spread scientific knowledge in a much more agile way, skipping the waiting time that normally takes in a verification process. This is one of the keys, along with the open culture, of Machine Learning that has allowed the field to develop at the rate it is going in recent years.

Keys to Simplify the Read of a Scientific Paper

Paper’s Structure

The structure is usually common in most papers but it can vary in the different fields of study. Generally, a paper is divided into different sections:

  • Abstract: A summary as a presentation of the content of the article.
  • Introduction: Presentation of the problem to be solved.
  • Related Work: Present the research line to which the paper belongs. Key to get related bibliography.
  • Materials and Methods: Explanation of how the problem has been solved. Here we will find most of the formulas, pseudo-codes, and details of the implementation.
  • Experiments and results: Section with graphs, visualizations, and tables
  • Conclusions: Summary of the results achieved as a result of the research
  • References: Of knowledge obtained from other pieces of work
  • Attached Documentation: In case there is one point that needs further explanations to be fully understood.

How to Read a Paper?

The first thing to be done is to read the abstract to assess if the topic is interesting to you and if you want to delve into the subject.

Then it will depend on how much time you want to invest in the subject or what level of mastery you have in it. If you do not know the field of study of the paper, you can focus on the related work, which is usually key to contextualize the subject matter, as well as the conclusions and results. Once a researcher already has a certain level of knowledge, they usually go straight to the experiments.

As an annex, the paper referenced below on how to read a paper is truly useful, it lays out a strategy on how to read papers that I personally have used a lot and has simplified greatly the process for me:

https://web.stanford.edu/class/cs244/papers/HowtoReadPaper.pdf

In summary, it proposes to read any given paper in three reading phases:

  1. First read: A quick overview of the abstract to assess if the paper is interesting to you
  2. Second read: This one is done pretty quickly, without reading formulas nor pseudocode and without trying to understand all the demonstrations of the experiments performed. Just reading the Introduction and the Conclusions.
  3. Third read: In which you are going to really delve into all the sections explained above and after which you should be able to reproduce what you have read.

Understand the Academic Point of View

Although this point is sometimes obvious, it is necessary to highlight it. We tend to think that papers, being scientific documents, are produced in a perfectly rigorous way, they follow agreed conventions and methodologies. Nothing could be further from the truth.

Being Machine Learning one of the most multidisciplinary scientific fields, as it feeds from Mathematics, Linguistics, Computer Science, Signal Processing… Each one of them has its unique set of methodologies. This means that in one paper a neural network is explained from its layer structure, in another paper through a signal processing algorithm, and in another through Bayesian probability formulas.

To fully comprehend a topic, normally it is necessary to analyze it from all its perspectives and if you want to learn more about a specific way of conceptualizing the problem (ie, Bayesian probability) you should review publications with shared magazines or shared conferences which would usually have a similar perspective.

Search for Opinions and Reviews About the Paper

Seek opinions on the paper and learn to be critical. When you read a paper, you should bear in mind that although it is a document that has passed some verification tests, this does not make it an error-proof document (especially when reading preprints).

Keep a critical spirit and ask yourself at all times if what you are reading is correct:

  • Is it methodologically okay?
  • Are the results well presented?
  • Do the graphs and visualizations follow good practices?
  • Does the paper solve the problem it presents?
  • Is it consistent between pages with the nomenclature used?
  • Is it advancing in the line of investigation? Or is it a more progressive advance that doesn’t have as much impact?

Many good and bad practices can be incurred when writing a paper. A very interesting resource is the OpenReview webpage, in which very good people make their reviews openly and attack works that have not been done with sufficient rigor and quality.

At this point, we should also highlight Yannic Kilcher‘s Youtube channel, which not only makes spectacular and immediate reviews of the most relevant papers that are being published but also usually complements them with personal evaluations and opinions as to why the papers may not be right methodologically (from which you can learn a lot).

Don’t (just) Read Papers

Don’t just read papers, be multimodal, go to different sources when learning about a topic that interests you.

Personally, I do not like papers, I like how they contribute to the growth of collective knowledge, but I do not think that they are the most appropriate means of spreading knowledge, they are difficult to digest and the reader has to make (normally) a great effort when it comes to really assimilate everything they propose.

In addition to the fact that most readers, who are not so specialized in that field of study, do not need such a deepening in these topics and surely there are many more suitable sources for them in that phase of their learning. An example of them is the articles that exist in the domain of Machine Learning and that I think they perform great pedagogical work and are pretty dynamic when presenting the content, since normally what they seek is that the reader can understand the insights in the most accessible way possible.

So, I invite you that if you are reading a paper and you do not understand fully what it presents, that you complement the experience with reading blogs on that topic. As there are many good alternatives today. For example, articles on Medium are of the highest quality, especially those published in Towards Data Science.

It can also be interesting to go to read discussions in forums on the subject, or if you are an expert, go to the GitHub repositories where these solutions are implemented. If you are interested in the latter, surely you are also interested in having a look at the page: PapersWithCode that lists publications that have their GitHub page associated with the code they refer to in the paper. All of these are good alternatives to expand knowledge.

In my personal opinion, a paper is a tool to transmit knowledge which is not updated to the new resources available in the 21st century. They have their place in an academic environment but for most Machine Learning practitioners they are not necessary and through the internet we find a lot of pedagogically excellent resources.

A case of success is the Distill.pub page that, apart from presenting the concepts with great clarity, consists of a multitude of interactive elements that allow any reader who accesses this page to access the knowledge in a much more accessible way. And we are not going to fool ourselves, this is very cool.

Finally, if you really want to delve into the content of papers, you can always form study groups or seminars where you go with a group of people who have studied and prepared the content of the paper to discuss it and ask questions about points that have not remained clear. Sharing your vision on a topic and talking about it is a great resource when it comes to learning about something. And most importantly, when one teaches, two learn.

Conclusions

Like always, I hope that you have enjoyed the post and that you are now armed with a good strategy to extract the maximum amount of knowledge of the papers you read (if you decide to do so).

If you liked this post then you can take a look at my other posts on Data Science and Machine Learning here.

If you want to learn more about Machine Learning, Data Science and Artificial Intelligence follow me on Medium and stay tuned for my next posts


Related Articles