I found out about the just concluded NIPS 2017 conference through a mention on my Twitter timeline. I’d heard this was a premier venue for presenting research in artificial intelligence, so I clicked through to the #NIPS2017 hashtag. The superlatives being thrown around to describe this year’s edition of the event roused me enough to watch some sessions for myself and check out a few demos as well. And boy, I was impressed. So herewith some notes. Also, isn’t it amazing how easily available all this research is even to people who have no direct relation with the AI community?
But first a quick question: What do you get when you ask a group of statistically minded people to organize a large conference? Well, detailed statistics about the conference itself, of course. The opening remarks by the conference chairs listed all the things that were new this year (a press conference, corporate after parties, contests for a bunch of ideas: chatbots, personalized medicine, learning to run, tricking AI systems and so on) and then used numbers to effectively convey how significant the event has become:
- 3000(!) papers submitted (an exponential increase) from more than 7000 people. Less than 700 were accepted (21%)
- 8000 people attending the conference (again a near doubling)
- 9 top level subject areas that break out into 156 low-level areas
- A corresponding increase in paper reviewers. Innovations like a mixed linear-integer software program to help with review assignment.
If you have time, don’t miss the part of the video (around the 15:25 mark) where they describe the steps they took to make the review process free from bias.
Naturally, the top tech companies of the world have been doing their homework ahead of NIPS. Google wrote a blog post which says they were planning to send more than 400 people! Amazon says they have 200 people, Microsoft and Facebook too have dedicated pages talking about their participation. Some slick videos from the after parties are also up. At the Intel after party they unveiled something called a “Neural Network Processor”! Nvidia’s after party also included a new processor, but more importantly it featured an orchestra playing AI generated music! Looking at the impressive corporate presence, one media report wrote about the “hiring wars” that were going on, and called presenters the new investment bankers! Ominous, I know 🙂
Since I find AI fascinating for the new and unexpected things it lets people do, I thought of taking a peek into the two pages (1, 2) of accepted demo applications. There is a “screen protector” app which figures out if someone is stealing a glance at your phone screen from behind you (don’t miss its funny video). There’s an image classifier app that learns quickly from your examples. Similar learning from examples is described for quality control in a food processing plant, in a project called Sensomind by Microsoft. Also, one of the best paper awards this year went to a Poker system that defeated top human players, and it has already been covered by the media.
As I suspected, the meat of the conference — namely the list of accepted papers — consists of long and intimidating titles that appear far removed from anything I might urgently care about. At the same time, I know I will end up hearing echoes of this list in mini gadgets around the house by the time the next holiday season rolls around, and so will hundreds of millions of people around the world. The host of the Intel after party pointed out how he was also a presenter a few years ago, and now manages a product at Intel that uses his work! Events like NIPS demonstrate the power of the academic-corporate nexus in the US. This smooth and rapid flow of advanced ideas to commercial products has made its corporations leaders on the global stage, and also helps supply its academic establishments with unrivaled talent.
Amidst all this self-congratulation, one dissenting voice came from Google scientist Ali Rahimi, who was awarded the “test of time” paper for work he did in 2007. His remarks during the acceptance speech caused a stir. He took a polemical shot at how research is being conducted in this field and told the people present that machine learning has become like alchemy! (video — start at 11:00 and feel free to skip the formulas). In his opinion, its practitioners, instead of applying tools on simple configurations and working their way up to more complex methods, have moved too quickly to higher levels without getting a clear understanding of the supporting layers. While this may be acceptable for harmless applications that are used for sharing photos or videos, the same attitude could lead to terrible consequences if carried over to critical products like cars or medical treatments, and might even have an adverse effect on the results of elections(!). To counter this, he urged the scientists to pause whenever they notice some mysterious behaviour while running their experiments, and then dig into the causes rather than quickly switch to a different technique. “Simple theorems and simple experiments” ought to become the building blocks of AI, in his opinion.
I also found detailed coverage of NIPS on the Insight Data Science blog.