Each day, more than 100 new computer science and machine learning papers are listed on arXiv. Though the works are not necessarily peer-reviewed before listing, this still is an enormous wealth of information. To get an impression, see the below chart for the growth of monthly submissions since 2009, taken from arXiv:

Doing the math, let’s assume that one needs 3 hours to read a paper from end to end, on average. At the numbers listed above, one would need 300 hours (or 12 days!) to read through them all. And that is just going through the papers of one day – the next day, we’d have to start anew; going through a similar number of publications again. Obviously, that’s not feasible, neither for experts nor for beginners.
Generally, as a beginner in Machine Learning, you are likely asking: do I need to read papers? And, given that there are so many, how can I do it at all? Here’s why and how!
Why you should read papers as a ML beginner
A paper is a lecture: to be accepted at top-tier ML conferences, publications need to be crisp in their writing. They include an introduction to the topic, a method section, results, and a summary. Altogether, the content of a paper is a (condensed) lecturing on a single, very narrow topic. For beginners, that is an excellent opportunity to get started in a field of their choice.
Well written papers introduce all the required terminology (either in the main section or expanded in the supplementary material) and categorize related works into a taxonomy. Thus, reading through a papers helps you scratch a mental map of the reasearch field. As you read more papers, you refine existing or add new areas to this mental map.
The process of reading and (unconscious) mental mapping helps you ask critical questions to the paper. Here, critical questions could be: where are the experimental details? Which augmentations were chosen? How has the data been normalized? Repeatedly going through this also translates to your coding practice: you avoid mistakes that you found others have done.

How to get into reading papers as a ML beginner
At the early stage, I recommend selecting a field of your interest. Fields could be computer vision, natural language processing, reinforcement learning, visualization techniques. Then, from your selected field, search papers published at top-tier peer reviewed conferences. In the ML field, these are: NIPS, ICLM, CVPR, ICLR, CVPR, ECML, among others. Alternatively, you can browse the top-tier journals, such as JMLR.
The peer-reviewed part is important. In peer-reviewing, researchers review your submitted manuscript; and in the ideal case – double-blind reviewing – you neither know the reviewers nor they know you. This process helps ensure that the paper adheres to certain quality standards, both in the actual content as well in the presentation (read: red thread throughout the paper) of the material.
After you have selected target venues, search for interesting papers. You might select them by their title, nice visualizations (examples that caught me to read a paper: CKA visualizations, loss landscapes), or by checking for the (non-)amount of mathematical expressions contained.
In your search, restrict yourself to publications that are 2 years or older. That restriction helps you lay a better foundation and won’t overwhelm you with too many newer advancements. Keep the hot, most-recent papers for later.
After you have collected a decent amount (5 to 20), start reading. You can read through the papers in any order, there does not need to be a chronology.
Expect that the first papers will overwhelm you, that is normal. For me, it took 3+ hours when I started seriously reading literature from my research field (continual learning: primer, scenarios, metrics). With practice, this has decreased to 1.5 hours.
Generally, it does not really matter how much you understand in the beginning; that you read them is what counts.
Conclusion
Beginners should not be scared by the growing number of machine learning papers published. As a beginner in machine learning, each paper is a valuable standalone lecture on self-selected topic. Reading through them helps you explore your field of interest better and hones your analytical thinking. To get started, simply select a ML subfield and pick not-too-recent (2 to 7 years old) papers.
Happy reading and learning!