We are living in an era of superlatives. Each year, month, week, new advancements in machine learning research are announced. The number of (ML) papers added to arXiv is growing equally fast. More than 11 000 papers have been added last October in the Computer Science Category.
Similarly, large machine learning conferences are seeing ever-growing number of submissions – so many in fact, that, to ensure a fair reviewing process, submitting authors are required to serve as reviewers for other submissions (called reciprocal reviewing).
Each paper possibly introduces new research results, a new method, new datasets or benchmarks. As a beginner in Machine Learning, it’s difficult to even get started: the amount of information is overwhelming. In a previous article, I argued that and why ML beginners should read papers. The quintessence is that good research papers are self-contained lectures that hone analytical thinking.
In this article, I give beginners ideas on how and where to find interesting papers to read, a point that I did not fully elaborate previously. Over 7 steps, I guide you through the possible process of finding and reading interesting papers.
Step 1: Determine your field of interest
Before you can find interesting papers, you should have an idea of the fields that interest you. For beginners, I recommend a dual approach: coding practice combined with research literature. In coding your own projects, you get direct hands-on practice and can test various fields of interest. A quick web or GitHub search for "machine learning starter project" gives multiple results (e.g., [this](https://machinelearningmastery.com/7-machine-learning-projects-beginners/) or this). Oftentimes, these articles are acccompanied by code. Skim through them and try projects that catch your attention.
Don’t overplan: doing anything usually is better than merely creating a perfect plan.
After you have tried several projects, you will have a rough understanding of whether a paper fits your topic or not. We thus can proceed to the next step.
Step 2: Find a ML conference
Leading ML researchers and research institutions publish at top-tier conferences. There are topic-specific as well as general ML conferences.
If you are generally interested in learning about ML, then one of the below conferences (respectively the papers published there) can be a good source of reading material:
- NeurIPS (also abbreviated as NIPS; Neural Information Processing Systems)
- ICML (International Conference on Machine Learning)
- ICLR (International Conference on Learning Representations)
- IJCAI (International Joint Conference on Artificial Intelligence)
If you are interested in computer vision projects and would like to learn more, than these conferences are a good starting point:
- CVPR (Computer Vision and Pattern Recognition)
- ICCV (International Conference on Computer Vision)
- ECCV (European Conference on Computer Vision)
If your interest is in the field of processing language and text data with neural networks, then these conferences are a good starting point:
- ACL (Annual Meeting of the Association for Computational Linguistics)
- EMNLP (Conference on Empirical Methods in Natural Language Processing)
- NAACL (Nations of the Americas Chapter of the Association for Computational Linguistics)
Lastly, for Research on audio or time series processing, the following conference could be a resource:
- ICASSP (International Conference on Acoustics, Speech and Signal Processing)
The above conferences are highly regarded in their respective fields. If your field of interest is not explicitly listed like, say, reinforcement learning, then look at the general conferences. There, you will likely also find papers for your interest, and not only at topic-specific venues. Also, there are many more conferences that I have not listed above, and I have not listed journals. Through your papers’ reference section, you naturally discover further venues. Lastly, implementing coding projects often also consists of implementing research algorithms. You can check the paper behind the algorithm to find out where it has been published.
If in doubt about the quality of a conference, you can enter the name on http://www.conferenceranks.com/ to get a first rating.
Before continuing to the next point, I’d like to emphasize that for ML beginners, I recommend to initially focus on peer-reviewed publications.
Why (and what is) peer-reviewing? Peer-reviewing is process where papers submitted to a conference are reviewed by other researchers from the same field*. It is often combined with a concept called "double-blind" reviewing: the paper is anonymized (meaning that identifying information is removed; see this paper for an example), and the researchers reviewing the paper are anonymous, too (see the comments here).
At the top-tier conferences, such a reviewing process ensures constant high-quality of the published material: papers adhere to scientific standards such as a fair comparison with related work, a good presentation of the content, a well-described method section.** This generally makes it easier for beginners to read papers.
Step 3: Search the conference proceedings and build a reading list
After step 1, you have identified your field of interest. After step 2, you either have found a topic-specific conference, or resort to the general conferences. In step 3, we are now going to find the actual papers. To make it concrete, I will use myself as an example.
- My research interest is continual learning (a topic I have written about [[here](https://medium.com/towards-data-science/continual-learning-a-primer-e328ed1d072f)](https://medium.com/towards-data-science/continual-learning-the-three-common-scenarios-e6c3260fe0cb), here, and here). Continual Learning is a general research area without a specific conference. I thus look at the last ICLR conference (2024 edition): https://iclr.cc/Conferences/2024:
- I then browse the list of accepted papers at https://iclr.cc/virtual/2024/papers.html?filter=titles by entering my search phrase "continual learning". This search gives me multiple results, visualized in below screenshot:
- Scrolling through the list, I mentally pre-filter the papers by title. Anything that sounds too mathematical, I happily skip. Of course, this will make me miss some research, but that’s fine: there’s plenty other research for me.
- Then, some title catches my attention: Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning
- By reading the abstract, I decide that the paper is interesting: I add it to my list of reading material. Reading the abstract is, however, only one way to find papers. Other ways could be: looking for the (non-)amount of mathematical equations, qualitative visualizations, or familiar authors.
To get a good set of starting papers, you can repeat the steps described with other conferences. Aim at selecting an initial set of 10 to 20 papers – you will find more later.
Step 4: Look for the older references
When I was at ICLR this year in May, I was listening to a talk at DeepMind’s booth. At some point, a listener from the audience remarked that "back then, people had more time and you notice this in the writing of the paper" (back then referring to pre 2020s).
As of writing this article, most ML research areas, at least when using neural networks, are fairly young. Thus, the deep learning foundations have been laid around a decade ago. Sometimes, these older research feels less rushed. It’s difficult to describe this aspect in writing, but once you have read several of the classics from your field, you might relate.
This older research is what we are now after: the older, often-cited papers. The respective older algorithms are well-studied; in some sense they are mature. To find them, look through your papers’ method and related work sections and search for methods that are cited and used as comparison again and again. In the field of continual learning, this are, among others, Elastic Weight Consolidation and Progressive Neural Networks. Also add the respective papers from your field to your reading list.
Step 5: Read through your list
This last step is quickly described: Read through the papers from your list, end-to-end. Expect it to take several hours per paper, at least in the beginning. With practice, the time required reduces. Also, use ChatGPT and similar tools to help you understand papers. You can ask them to explain a paragraph or check if you understood something correctly.
As you read through the papers, mark any references that sound interesting. They will be the foundation for your next reading list.
Step 6: Read a survey
Survey papers are the ideal way to get an overview about a research field. Continuing with my example from before, a recent survey for my field of interest would be this one.
However, I would not recommend reading a survey as the first paper, at least if you are a beginner in the field. The reason is that surveys are naturally very dense: They are covering a research area in great detail, and often the reference sections are multiple pages long! Reading through such a survey might be overwhelming for beginners. You will benefit more from such papers when you already have a rough mental map from your field. At that point, a survey can help you refine your map.
If you have prior experience in another research field (say, in computer vision), then reading a survey to learn about a novel field (say, natural language processing) is easier.
Step 7: Finish your list
That will never happen. Once you got hooked, you will always have something to read on your list. It’s like the Steam Pile of Shame: always adding new papers, but never being able to read them all.
Happy reading!
*That’s the ideal case. In practice, when a paper is reviewed by 4 reviewers, it can happen that one of the reviewers is from a different discipline. ** Again, the ideal case. There always are some fluctuations in quality. The system is made by imperfect people for imperfect people.
