I was just at NeurIPS (gave a spotlight talk in the session on online learning). TBH, NeurIPS gas grown too big for it's own good; this year there were 14,000 attendees. It's so big that they had to restrict access to some of the poster sessions due to overcrowding. It's become very difficult for junior researchers to find and talk to senior people, due to the sheer size of it all; honestly, it feels like an overcrowded science fair crossed with an industry expo. I much prefer AISTATS, which I attended this year in Okinawa, Japan. Much smaller, more accessible, and with more focus on disseminating scientific ideas, as it should be.
Probably also too many people are focused on a solution, and not problems for specific domains. Learning more about deep learning is cool, but in practice, most problems in domains need domain expertise combined with more mundane models. This is particularly true when you need to rally domain experts help - folks able to understand logistic regression more easily than a complex deep learning model.
NeurIPS should be about fundamental science not about applied science and engineering solutions for specific domains.
Science vs applications would be actually a good way to divide NeurIPS into two conferences to get focus back: NeurIPS and Applied NeurIPS. One wold present science, the other tweaks, tricks and methods to solve everyday problems.
Papers like "Partially Encrypted Deep Learning using Functional Encryption", "Dancing to Music", "Differentiable Cloth Simulation for Inverse Problems" would go go the engineering conference for the "most people".
Well machine learning is applied science. But inside there are theoretical work (like encrypted machine learning or optimization theory) and practical work (like text2speech, or other tasks)
Disagree. Theory is always interested in building better, simpler protocols, asking questions about lower bounds and possibility. See almost all of fundamental crypto research.
I had a workshop spotlight. Those sessions were generally a lot calmer.
A big problem was cramming hundreds of posters in one large space, with a relatively poor search experience (posters were kind of arbitrarily organised with some categories only having three or four entries). If they'd split them over both exhibition centres (or multiple rooms/floors), it might have calmed the crowds. Also since it was hard to know exactly where to look, there was a lot of wandering around.
There's also the heavy corporate side where people spent the evenings going to not-so-secret invite-only parties (lots of FOMO going on). Lots of jobs going, but ridiculous amount of competition now.
Despite that it was pretty friendly, and I had good results emailing people I wanted to chat to, and going to the socials.
Really the field needs to start publishing in journals again, and not obsessing over one or two mega-conferences each year. In other domains, there are huge congress events, but they're much more informal (eg EGU, 17k people).
I completely agree. I have been on ICCV this year (only 7500 participants) and found it really difficult to find people who I would like to talk to (e.g. because they had a nice presentation). The number of submissions has doubled since the last time (two years ago) [0]. It is just a constant overflow of information and the conference (including workshops, which were pretty good) lasted seven days.
I think these ML/DL conferences need to be split up and specialize more to manage the interest.
Perhaps I'll provide a contrary view. I was also at Neurips, and I'm also a junior researcher. Yes, it was massive, and the poster sessions were crowded.
However, I was still able to talk to everybody I wanted to talk to - often for long periods of time. Many of them I sent emails to, and I was able to talk with everybody I wanted to (except 1), many who would be considered senior researchers.
Disseminating scientific ideas is really only one part of what a conference is meant to do. The other part is networking, and that was extremely valuable for me.
Which bring so much software poverty...
So many tasks have a state of the art accuracy only on tensorflow or only on pytorch.
So now someone that actually care about accuracy must learn both frameworks instead of one but in practice, he will just not use the state of the art when not available which is just plain sad.
i understand why there would be differences wrt performance, ease of deployment, etc, but why would there be significant differences in model accuracy between tf and pyt? a matrix multiply is a matrix multiply, regardless of implementation...
The matrix multiplications should be the same (down to floating point accuracy limitations) but there might be very slight differences in how things like random dropout or stochastic gradient descent work in one framework versus the other.
I didn't even realize the conference was happening until my gf complained about the attendants of some ai conf downtown. Didn't quite realize how big it was.
Which? Her technical talk on how humans learn [1], or her keynote address on sexual harassment and the #metoo movement? As far as I know, only the latter got a standing ovation.
The same talk [1] covers both topics. Part of her point is that they're related via people forming false beliefs about accusations of harassment (particularly men believing that they're at much greater risk of being accused than they actually are).
