A Statistical Tour of NIPS 2017

The Conference on Neural Information Processing Systems (NIPS) was first held in Denver in December 1986. The leading annual gathering of its kind, NIPS has since visited a number of Canadian and Spanish cities and will again head up to Montréal in 2018. According to official statistics, NIPS 2017 in Long Beach, California was the most popular yet, attracting over 8,000 registered attendees, up 2,000 from last year.

Sponsorship

Industry is playing an increasing role in academic conferences. Corporate sponsorship and paper submissions from corporate labs are now a norm. The following information is based on sponsor contributions as listed on the NIPS official website.

The 31st NIPS attracted 84 sponsors (totaling US$1.76 million), an increase of 31.5% from the previous year (64 sponsors totaling US$840,000). In 2017, NIPS added a US$80,000 “Diamond” sponsorship category, while the Platinum sponsorship threshold rose from US$25,000 to US$40,000.

Soaring fees have not dampened sponsors’ enthusiasm. Five Diamond sponsors contributed US$400,000 to the conference this year. The more affordable Gold and Silver category sponsors also increased dramatically. Meanwhile, entry-level Bronze sponsors fell 70% from last year, the dip likely related to the committee’s more rigorous sponsorship screening process.

Half of the corporate sponsors were from the tech industry, followed by around 28% from the financial industry (banking, trading, and financial services). Our on-site research suggested that finance companies largely participated for recruitment purposes and to stay abreast the latest trends in machine intelligence. However, few recruiting companies were able to provide clear and concrete hiring objectives or job descriptions, instead they gave vague responses, for example that candidates should simply be “third-year PhD students,” or “able to analyze big data.”

Submissions

There were nine top-level subjects and 156 subareas at this year’s conference. The latter witnessed a 150% increase from last year, reflecting growing research diversification. The hottest subareas were: 1) algorithms; 2) deep learning; and 3) application.

We extracted technical keywords from all presentations and poster titles. Buzzwords such as “learning,” “deep,” and “neural networks” topped the list; with related terms such as “stochastic method,” “optimization,” “Bayesian,” “reinforcement learning,” “adversarial,” “inference,” and “Gaussian process” also frequently showing up. Interestingly, although “reinforcement learning” and “data” were smaller subareas, these terms appeared much more frequently in paper titles.

Publishing

According to statistics compiled by Infinia ML CEO Robbie Allen, the most active paper-submitting academic institutions were Carnegie Mellon University, the Massachusetts Institute of Technology, and Stanford University; while corporate paper submissions were led by Google, Microsoft, and IBM. Interested readers can pursue Allen’s breakdown analysis here.

Participation

According to sample survey of close to 100 conference attendees, we gathered that 77% percent came to NIPS to learn about AI trends, 15% for networking, and 8% for job opportunities.

Spotlights

Some NIPS 2017 events were streamed live on Facebook (excluding Symposiums and Workshops). Most of the videos had views in the hundreds one week after the conference ended. Among the more popular videos was Yann LeCun’s “Geometric Deep Learning on Graphs and Manifolds,” which received 8,300 views. A video combining speeches by Deep Genomics Founder Brendan Frey and Ali Rahimi’s “Machine Learning has become Alchemy” received 1,900 views. As with all contemporary events, social media popularity is a key factor in content distribution.

Based on on-site session attendance data extracted from the official NIPS phone App, we can see that aside from the poster sessions, Hall A’s tutorial attracted the most participants for the following keynotes: “Deep Learning: Practice and Trends,” “Deep Probabilistic Modelling with Gaussian Processes,” and “Geometric Deep Learning on Graphs and Manifolds.”