A Blog by Jonathan Low


Dec 14, 2017

What Goes On At the World's Biggest, Most Prestigious Artificial Intelligence Conference

Methodology debates and corporate recruiting. Not necessarily in that order. JL

Dave Gershgorn reports in Quartz:

Fights between AI’s most visible figures dominated: for engineers to understand more about the techniques they’re using - “Make Machine Learning Rigorous Again,”-  while others argued AI’s accuracy via open-sourced code and borrowed techniques is proof you don’t need to understand every bit of your algorithm. “We had really good ideas in the ’90s. And they are the same ideas that we see now. But now they work.”
Look into the sea of computer scientists shaping the future of artificial intelligence and the first thing you’ll notice is all the backpacks. Expensive and branded, the backpacks mark the mostly male 20-and-30-somethings as members of the tribes of Google, Nvidia, Facebook, and Microsoft.
These researchers gathered last week at the largest and most influential AI conference, Neural Information Processing Systems. This is the conference’s 40th year, and its most-attended, at 7,229 registrations (only 17.1% were women). The people at the conference are the kind who need to carry computers with them all the time—pecking out presentation slides minutes before jumping on stage, training their algorithms up to the last-minute to eke out optimal results, checking the latest academic mudslinging on Twitter.
But the researchers who weren’t wearing backpacks were the talk of this year’s conference. They’re the old guard, including some who have attended the conference for more than 10 years. After years in academic obscurity, they’ve largely been recruited into senior positions at the companies that hand out pricey backpacks. In between presenting the latest work from the AI labs they lead, the typically-reserved corporate department heads traded barbs and showed up at talks to debate each others.
DeepMind CEO Demis Hassabis, 41, whose team dropped new research showing that its world champion Go-playing AI had also mastered chess and shogi, stepped up to a microphone after a talk to rebut comments from NYU professor and ex-Uber AI guru Gary Marcus, 47. Marcus was critical of DeepMind’s AlphaZero, saying the company had done some “hacks” to get the results it presented. Facebook’s AI head Yann LeCun, 57, wrote to his 60,000 Facebook followers to dispute senior Googler Ali Rahimi’s call for more rigor in the field. The fights between AI’s most visible figures dominated the conference’s narrative. Rahimi’s call for engineers to understand more about the techniques they’re using was echoed in cries of “Make Machine Learning Rigorous Again,” while others argued that AI’s skyrocketing accuracy via open-sourced code and borrowed techniques is proof you don’t need to understand every bit of your algorithm.
The first four days of the conference also served as a feeding frenzy for companies looking to recruit students and poach talent from rivals. Coders that can work with machine learning algorithms are among the most sought-after in Silicon Valley, and those presenting papers on new research at NIPS are top prizes. The proof is in the salaries: A New York Times report found starting salaries for Ph.D level researchers could range from $300,000 to $500,000.
Conference attendees who spoke to Quartz said companies often don’t even wait for students to graduate from Ph.D programs, offering packages to those who have shown the ability to do cutting-edge research alongside coursework.
Recruiters constructed huge booths in the Long Beach Convention Center’s central hall, some vetting resumes of potential new hires on the spot. An IBM representative arranged an interview with HR after glancing at a CV, while an Uber recruiter said the company is hiring for “literally everything in machine learning.”
Companies also lured talent with lavish parties; rapper Flo Rida performed at Intel’s party, while a live jazz band performed for Facebookers as they enjoyed an “unparalleled exclusive view of the private jet runway at the award winning Long Beach Airport,” according to the invite. NIPS wasn’t always so corporate, and some attendees pined for the good old days. UC Berkeley professor Ben Recht wore a shirt that read “Corporate conferences still suck,” as he accepted the “Test of Time” award with Rahimi for a paper they presented at NIPS 2006. The shirt, an homage to Kurt Cobain’s “Corporate magazines still suck” shirt pictured on a 1992 cover of Rolling Stone, lit up the #NIPS2017 hashtag during the presentation. Risto Miikkulainen, a professor at University of Texas at Austin and VP of research at AI startup Sentient Technologies, attended NIPS in the late ’80s and ’90s when the conference was still a few hundred scientists working in an obscure field. He stopped coming to the conference in the early aughts, the last AI winter when funds for the technology dried up, but started attending again last year since the field has begun to reheat.
Miikkulainen is in a situation similar to those whose names were whispered between sessions and at bars: He’s been working on neural networks, the technology responsible for the current AI boom and only found to be viable in 2012, since before it attracted big dollars.
“This feels very different,” Miikkulainen says of this year’s conference. “We had really good ideas in the ’90s. And as a matter of fact, they are the same ideas that we see now. But now they work.”


Post a Comment