NeurIPS 2021 Showcases AI and ML Research: The 35th edition of NeurIPS concluded last week, offering an overview of the state of AI and ML research and which companies, universities and countries are leading the way. Of the 2,344 accepted papers, U.S. researchers led the pack with 1,431, well ahead of China’s 411 and the UK’s 268. Among private industry-affiliated papers, Alphabet subsidiaries took first (Google, 177 accepted papers) and third place (Deepmind, 81), with Microsoft in second (116 accepted papers). University-affiliated papers, meanwhile, saw U.S. institutions take four of the top five spots — China’s Tsinghua University rounded out a top five led by MIT, Stanford, Carnegie Mellon, and UC Berkeley. Six papers took home Outstanding Paper Awards, while two were recognized with the new Datasets and Benchmarks Best Paper Awards. NeurIPS also gives out a Test of Time Award, which recognizes older papers that have a continuing impact on the field. This year’s recipient was Online Learning for Latent Dirichlet Allocation, published in 2010.
- More: What was hot at this year’s NeurIPS AI conference | Inconsistency in Conference Peer Review: Revisiting the 2014 NeurIPS Experiment
- German regulators approved Mercedes-Benz’s Drive Pilot automated driving system, making it the world’s first commercially available level 3 self-driving system. Level 3 is the first of SAE International’s six automation levels in which the human driver can (under approved circumstances) take their eyes off the road while the car drives itself. Drive Pilot, which can hit German roads starting next year, is approved to operate only on specific premapped, geofenced highways at speeds up to 60 km/h (regulators have not yet decided whether to clear the system for speeds up to 130) — Mercedes-Benz says more than 13,000 km of roads fit those conditions in Germany.
- UK-based DeepMind introduced RETRO (the Retrieval-Enhanced Transformer), a language model that performs similarly to other large language models at only a fraction of their size. Taking inspiration from the human brain’s use of dedicated memory mechanisms, RETRO has a database of two trillion text passages that it uses like a reference sheet to compare and improve its text generation. It needs only seven billion parameters to match the performance of another 280 billion-parameter DeepMind language model. As the computational demands of large language models become unsustainable, finding ways to make AI systems more efficient will be essential.
- More: DeepMind AI tackles one of chemistry’s most valuable techniques | DeepMind’s AI helps untangle the mathematics of knots
Senate Passes NDAA, On to Biden for Signature: Last week, the Senate voted 88-11 to approve the FY2022 National Defense Authorization Act. The $768 billion bill (summary available here) contains a number of provisions related to AI and emerging technology, including:
- Sec. 217 requires the DOD to establish a national network for microelectronics R&D. The FY2021 NDAA had authorized such a network as part of the CHIPS for America Act; the new bill mandates it.
- Sec. 226 requires the DOD to “review the potential applications of artificial intelligence and digital technology to the platforms, processes, and operations” of the military and to establish performance metrics for AI adoption efforts.
- Sec. 228 requires the DOD to establish “executive education activities on emerging technologies” for select officers and civilian staff.
- Sec. 232 authorizes the Pentagon to carry out a pilot program that would establish data repositories containing DOD data. Those repositories would then be accessible to certain private and public organizations to aid in AI and ML development for the DOD.
- Sec. 247 requires the DOD to provide regular reports and briefings to Congress on its implementation of recommendations made by the National Security Commission on Artificial Intelligence in its final report.
- Sec. 851 modifies a section of the FY2021 NDAA that had prohibited the acquisition of printed circuit boards from North Korea, China, Russia or Iran. The prohibition had been set to take effect in 2023, but this provision delays it until January 1, 2027.
- Sec. 1251 directs the DOD to conduct a comparative analysis of U.S. and Chinese modernization efforts across five critical technologies: directed energy, hypersonics, biotechnologies, quantum science and cyberspace capabilities.
Biden Administration Ramps Up Crackdown on Chinese Companies: Over the past month, the Commerce Department has added dozens of Chinese companies to its Entity List — restricting their access to U.S. exports without government approval — while the Treasury Department blacklisted several more for their involvement in the oppression of China’s Uyghur population — barring U.S. citizens and firms from investing in those companies. Among the firms added to the Treasury blacklist are the drone company DJI Technology Co., four companies involved in developing facial recognition technology, and the AI giant SenseTime. SenseTime’s addition caused the firm to temporarily postpone its $770 million Hong Kong IPO, but the company chose earlier this week to relaunch it. According to The Financial Times, the IPO’s resurrection was made possible by hundreds of millions dollars of financial backing from Chinese state-backed investors, which will help SenseTime replace the investors it stands to lose post-blacklisting.
Job Openings and New Funding
CSET closes out the year with the happy news of a $250,000 grant from the Patrick J. McGovern Foundation to support the CyberAI project, particularly to highlight current barriers to cybersecurity and AI education at the high school level.
And we have more exciting news — we’re hiring! Please apply or share the role below with candidates in your network:
- AI Research Subgrant (AIRS) Program Director: CSET’s AIRS program will promote the exploration of foundational technical topics that relate to the potential national security implications of AI over the long term via research subgrants. The Director of AIRS will manage all technical, programmatic, and financial aspects of the new AIRS program.
What’s New at CSET
- Wisdom of the Crowd as Arbiter of Expert Disagreement by Michael Page
- Making AI Work for Cyber Defense: The Accuracy-Robustness Tradeoff by Wyatt Hoffman
- AI and the Future of Disinformation Campaigns: Part 2: A Threat Model by Katerina Sedova, Christine McNeill, Aurora Johnson, Aditi Joshi and Ido Wulkan (Read Part 1)
- Foreign Affairs: Soon, the Hackers Won’t Be Human by John Bansemer
- OECD — The AI Wonk blog: Testing Frameworks for the Classification of AI Systems by Catherine Aiken
- The Diplomat: Will China’s Regulatory ‘Great Wall’ Hamper AI Ambitions? by Ellen Lu and Ryan Fedasiuk
- SpyTalk Podcast: Smashed Plots and Then Some featuring Anna Puglisi
- CSET: Data Snapshot: Exploring Supercomputing with the Map of Science by Autumn Toney
Foretell has launched a new project that combines expert and crowd judgment. You can read more about the experts’ views, including how they think trends like China’s military aggression, political polarization, and the strength of the tech sector affect the DOD-Silicon Valley relationship. See all 20 forecast questions associated with this project here, and read more about the subject in Research Fellow Michael Page’s new issue brief, Wisdom of the Crowd as Arbiter of Expert Disagreement.
- On December 16, CSET’s webinar Deconstructing China’s Vision for the Future of Warfare featured a conversation between CSET Research Analyst Ryan Fedasiuk and the inaugural director of the DOD’s Joint Artificial Intelligence Center, Lieutenant General (retired) Jack Shanahan, about Chinese military progress in AI and its implications for the United States.
- The Washington Post: An Ellen Nakashima and Jeanne Whalen article about the Biden administration’s approach to Chinese companies with PLA ties cited Ryan Fedasiuk, Jennifer Melot and Ben Murphy’s October report, Harnessed Lightning: How the Chinese Military is Adopting Artificial Intelligence.
- The Wall Street Journal: Harnessed Lightning also earned a citation in a Liza Lin and Jing Yang piece about SenseTime’s postponed IPO.
- Marketplace: CSET Director of Biotechnology Programs and Senior Fellow Anna Puglisi spoke with Sabri Ben-Achour for a report about Chinese state-sponsored industrial espionage.
- Newsweek: Research Fellow Margarita Konaev discussed Turkish UAV development with Tom O’Connor for a recent story.
- The National: Konaev also spoke with Joyce Karam about the U.S. military’s shifting technical focus as it pivots towards the Asia-Pacific region.
- VentureBeat: A Kyle Wiggers article about the dangers of AI-enabled propaganda-as-a-service cited the CSET report Truth, Lies, and Automation: How Language Models Could Change Disinformation by Ben Buchanan, Andrew Lohn, Micah Musser and Katerina Sedova.
- The Wire China: Anastasiia Carrier drew on the August data brief China is Fast Outpacing U.S. STEM PhD Growth by Remco Zwetsloot, Jack Corrigan, Emily Weinstein, Dahlia Peterson, Diana Gehlhaus and Ryan Fedasiuk for a report on China’s high skill labor shortage.
What We’re Reading
Paper: The Great Tech Rivalry: China vs the U.S., Graham Allison, Kevin Klyman, Karina Barbesino and Hugo Yen, Belfer Center for Science and International Affairs (December 2021)
Article: Filling gaps in trustworthy development of AI, by Shahar Avin et al., Science (December 2021)
Article: Ethical and social risks of harm from Language Models by Laura Weidinger et al., arXiv (December 2021)
- January 20: CSET and Stanford HAI Webinar, Strengthening the Technical Foundations of U.S. Security, featuring CSET Senior Fellow Andrew Lohn, Stanford HAI Director of Policy Russell Wald and Stanford HAI Postdoctoral Fellow Jeff Ding
What else is going on? Suggest stories, documents to translate & upcoming events here.
policy.ai will be off for the holidays, but don’t worry — we’ll be back on January 13 with all the latest in AI, emerging tech and security policy!