NeurIPS

NeurIPS is a large gathering of researchers who teach computers to learn from examples. It takes place once a year, usually in December, and lasts about a week. The name stands for “Conference on Neural Information Processing Systems.” Anyone who wants to present something there submits a written academic paper beforehand. Other experts review this text anonymously and decide whether it is good enough. Only a quarter to a third of submitted papers are accepted.

The pacesetter of AI research

In many sciences, important results appear in academic journals. In AI research, it’s different: here, conferences count for more. NeurIPS is the most well-known of these, alongside ICML and ICLR. A paper accepted at NeurIPS is considered solid proof of a career.

The conference is also an early warning system. Many methods that are now embedded in language models were presented there years earlier. A famous example is the 2014 paper on so-called Generative Adversarial Networks, in which two programs compete against each other to generate images. Anyone who wants to know what companies will be making money from in three years reads today’s conference papers.

Economically, NeurIPS is also a huge job market. Google, Meta, OpenAI, and Nvidia maintain booths there and recruit PhD students. Attendance rose from a few hundred in the 1990s to over 15,000 at times. Tickets have repeatedly sold out within minutes.

From submitted paper to talk

The process follows the same pattern every year. In spring, research groups submit their papers, usually eight to ten pages long. Then the review process begins: three to five experts read the text and write critiques. They don’t know the names of the authors, and the authors don’t know the names of the reviewers. This double-blind process is meant to prevent famous names from being favored.

Afterward, the authors are allowed to respond to the critiques and clear up misunderstandings. Only then is the decision on acceptance or rejection made. Almost all accepted papers are presented as posters: you stand next to a large poster board and explain your findings to anyone who stops by. Only a few particularly outstanding papers get a talk in front of a large audience.

A well-known problem is the randomness of this selection process. In an experiment in 2014, part of the submissions were reviewed twice, by two independent groups. The two groups disagreed on more than half of the accepted papers. A rejection therefore does not necessarily mean that a paper is bad.

NeurIPS in news and headlines

In business news, NeurIPS usually appears as a time reference. Sentences like “the company presented the method at NeurIPS” frame a piece of news as serious research, not as advertising. Stock movements at chipmakers are also sometimes attributed to announcements made at this conference.

You also encounter the name on servers like arXiv, where academic papers are available to read for free. Below the title, there is often a note like “NeurIPS 2023.” This is an indicator of quality: the text has survived peer review. Without such a note, anyone can upload anything, unchecked.

A side note explains a common confusion. Until 2018, the conference was simply called NIPS. Because this abbreviation also has a lewd meaning in English, it was changed to NeurIPS. Older sources therefore refer to NIPS, but mean exactly the same event.