[Rough Draft] Going to NeurIPS is Weird

In the course of my PhD studies, I’ve published at, reviewed for, and attended a number of large AI conferences. Among these have been ICML, KDD and NeurIPS. In general, while I don’t want to make it my life’s work, working as a researcher in academia has been a largely positive experience. It has taught me humility, perseverance and showed me just how difficult meaningful research actually is. It also made me a much better public speaker, software engineer and teacher.

However, there has also been a fair amount of disillusionment about the research process, especially the current system of research publishing and funding. Smarter people with much more experience have written about this problem at length, so I do not want to retread these paths on my blog. A small part, which I don’t see mentioned that frequently, has been with AI conferences in particular.

For one, I believe that the current scale of the top-level AI conferences is detrimental to (public?) research in the field. Of course, there are other, more immediate issues, such as the lack of funding for universities to participate in researching foundation models or the general political climate, with the US engaging in power-plays to protect its leading position¹ by enacting tarrifs and export controls². However, when conferences publish hundreds or thousands of papers, it becomes almost impossible for researchers to exchange ideas and connect to people working in similar areas. The poster sessions become hectic and the presence of many industry folks asking about practical applications of the research significantly worsens the signal-to-noise ratio for those presenting a poster.

I believe the researchers are not at fault in this whole deal, as publishing at these conferences is the de-facto way of opening doors for grant applications and obtaining tenure.

In closing, I want to point out one thing I have perceived as strongly positive compared to the conferences in more mathematical domains: The advocacy and networking groups such as Queer in AI, latinx in AI and wiml, among many others. These groups do great work in ensuring that people from marginalized communities feel welcomed at these conferences and that they have the necessary accomodations and a support system to turn towards, if they experience harassment. Of course, the large scale in some sense also helps with establishing such groups and ensuring that there is sufficient funding and space to accomodate workshops and other initiatives, but I do believe that other compute science domains would greatly benefit from such volunteer-led groups existing.