Could an AI write scientific papers of the future, and possibly win the Nobel Prize? This was the focus of Yolanda Gil’s Presidential Address at the Association for the Advancement of Artificial Intelligence (AAAI) 2020 Conference (AAAI-20) in New York, Feb. 7-12.
Gil, a research professor of computer science and a principal scientist at the USC Information Sciences Institute, believes that universities play a critical role in the advancement of science and discovery via artificial intelligence. Gil was elected president of AAAI, the largest international scientific society in AI, in 2016, when interest in AI was just taking off. She is one of the principal authors of the country’s AI roadmap, which looks into the future of AI research in the next 20 years.
Gil says industry tends to work on short-term opportunities, but universities have a duty to prepare for the long term by studying challenging problems, such as reducing the time it takes to make scientific discoveries. She told the audience of about 4,000 at AAAI-20: “If you look at the problems that we face: understanding the brain, understanding the planet, the universe, our environment…we need AI to help us push the frontiers of science.” Structurally, multi-university centers are needed for AI research in this new era, one that is more experimental in nature, Gil says.
“We need AI to help us push the frontiers of science.” Yolanda Gil.
She explained to the AAAI audience: “Human limitations curb scientific progress…when we write papers, and when scientists look at the world, sometimes there are errors—there are biases.”
AI, says Gil, presents an opportunity to eliminate biases by searching systematically through alternative hypotheses to eliminate errors by incorporating constraints and other forms of scientific knowledge. At the same time, she acknowledges that human error often leads to discoveries, as was the case in the discovery of penicillin. So, while AI could make discoveries more efficiently, humans may make complementary discoveries, which are different in nature.
“The combination of both AI and humans doing research has the potential to lead to fundamentally new kinds of discoveries,” she said. Gil states that there is not just one pathway for AI research. She added: “diversity and breadth are very important, particularly in academia.”
She also stressed the importance of universities focusing on the use of AI for areas of societal importance that are not addressed by industry, such as scientific research, education, healthcare, innovation, and social justice. At the conference, she announced a million-dollar annual prize for work on “AI for the benefit of humanity.”
With only significant recognitions, such as the Nobel Prize, offering that level of award, AAAI hopes to raise awareness and encourage work on such AI applications.
A forum for different thoughts and perspectives
USC Viterbi was well represented at the AAAI-20 conference. Fei Sha, the Zohrab A. Kaprielian Fellow in Engineering and associate professor of computer science and biological sciences, was one of two program co-chairs in the conference, which drew a record-high number of about 8,000 full-length submissions.
Sha talked about the conference’s role in broadening the community of researchers in various subfields of AI, from the natural sciences and engineering to social sciences and medicine, thus providing a forum to reflect on the history and progress of artificial intelligence and contemplate directions for its future. Some of the emerging topics were AI for social good, AI for impact and AI for medicine.
Sha says that the organizing idea of the conference was, “to present a forum for different thoughts and perspectives, which can synergize and spark new ideas.” Keynote speakers included three Turing Award winners, a Nobel laureate in economics, and a number of other notable experts.
For example, the history panel included Garry Kasparov, the chess champion, who was beaten by DeepBlue, as well as researchers representing IBM Watson, the AlphaGo team from DeepMind, and RoboCup, who discussed how “games can advance AI research.” Sha wondered, “whether we can actually design games that humans can win over AI game players.” A discussion ensued about how human and AI co-exist.
With a background in statistical machine learning, one of the major forces advancing today’s AI, Sha believes that the breadth of the research problems at AAAI 2020, and their potential impacts, is immense. Particularly, he said, in learning paradigms of complex decision-making in uncertain environments, beyond perception and pattern recognition, where data for learning are limited.
USC at AAAI-20: Publications
According to data analyzed by the Microsoft Academic project, USC was the third top publishing institution at AAAI, after CMU and MIT. This year, USC participated with 15 publications on areas as diverse as object removal for autonomous driving, idle time optimization and human-machine collaboration. A list of papers presented is below:
MIPaaL: Mixed Integer Program as a Layer
Aaron Ferber (USC); Bryan Wilder (Harvard University); Bistra Dilkina (USC); Milind Tambe (Harvard University)
Modeling Dialogues with Hashcode Representations: A Nonparametric Approach
Sahil Garg (USC); Irina Rish (University of Montreal); Guillermo Cecchi (IBM); Palash Goyal (USC Information Sciences Institute); Shuyang Gao (USC Information Sciences Institute); Sarik Ghazarian (USC Information Sciences Institute); Greg Ver Steeg (USC Information Sciences Institute); Aram Galstyan (USC Information Sciences Institute)
Invariant Representations through Adversarial Forgetting
Ayush Jaiswal (USC Information Sciences Institute); Daniel Moyer (USC Information Sciences Institute); Greg Ver Steeg (USC Information Sciences Institute); Wael Abd-Almageed (USC Information Sciences Institute); Prem Natarajan (USC Information Sciences Institute)
Communication, Distortion, and Randomness in Metric Voting
David Kempe (USC)
An Analysis Framework for Metric Voting based on LP Duality
David Kempe (USC)
Idle Time Optimization for Target Assignment and Path Finding in Sortation Centers
Ngai Meng Kou (Alibaba); Cheng Peng (Alibaba); Hang Ma (Simon Fraser University); T. K. Satish Kumar (USC); Sven Koenig (USC)
Predictive Engagement: An Efficient Metric for Automatic Evaluation of Open-Domain Dialogue Systems
Sarik Ghazarian (USC Information Sciences Institute); Weischedel Ralph (USC Information Sciences Institute); Aram Galstyan (USC Information Sciences Institute); Nanyun Peng (USC Information Sciences Institute)
Generative Attention Networks for Multi-Agent Behavioral Modeling
Guangyu Li (USC); Bo Jiang (DiDi AI Labs, Didi Chuxing); Hao Zhu (Peking University); Zhengping Che (DiDi AI Labs, Didi Chuxing); Yan Liu (USC)
End-to-End Game-Focused Learning of Adversary Behavior in Security Games
Andrew Perrault (Harvard University); Bryan Wilder (Harvard University); Eric Ewing (USC); Aditya Mate (Harvard University); Bistra Dilkina (USC); Milind Tambe (Harvard University)
AutoRemover: Automatic Object Removal for Autonomous Driving Videos
Rong Zhang (Zhejiang University); Wei Li (Baidu); Peng Wang (Baidu USA LLC.); Chenye Guan (Baidu); Jin Fang (Baidu); Yuhang Song (USC); Jinhui Yu (Zhejiang University); Baoquan Chen (Peking University); Weiwei Xu (Zhejiang University); Yang Ruigang (Baidu)
MetaLight: Value-based Meta-reinforcement Learning for Traffic Signal ControlXinshi Zang (Shanghai Jiao Tong University); Huaxiu Yao (Pennsylvania State University); Guanjie Zheng (Pennsylvania State University); Nan Xu (USC); Kai Xu (Shanghai Tianrang Intelligent Technology Co., Ltd); Zhenhui (Jessie) Li (Penn State University)
Toward A Thousand Lights: Decentralized Deep Reinforcement Learning for Large-Scale Traffic Signal Control
Chacha Chen (Pennsylvania State University); Hua Wei (Pennsylvania State University); Nan Xu (USC); Guanjie Zheng (Pennsylvania State University); Ming Yang (Shanghai Tianrang Intelligent Technology Co., Ltd); Yuanhao Xiong (Zhejiang University); Kai Xu (Shanghai Tianrang Intelligent Technology Co., Ltd); Zhenhui (Jessie) Li (Penn State University)
To Signal or Not To Signal: Exploiting Uncertain Real-Time Information in Signaling Games for Security and Sustainability
Elizabeth Bondi (Harvard University); Hoon Oh (Carnegie Mellon University); Haifeng Xu (University of Virginia); Fei Fang (Carnegie Mellon University); Bistra Dilkina (USC); Milind Tambe (Harvard University)
Human-Machine Collaboration for Fast Land Cover Mapping
Caleb Robinson (Georgia Institute of Technology); Anthony Ortiz (University of Texas at El Paso); Nikolay Malkin (Yale University); Blake Elias (Microsoft); Andi Peng (Microsoft); Dan Morris (Microsoft); Bistra Dilkina (USC); Nebojsa Jojic (Microsoft Research)
Recent Directions in Heuristic Research (tutorial)
Ariel Felner (Ben-Gurion University), Sven Koenig (USC), Nathan Sturtevant (University of Alberta) and Daniel Harabor (Monash University)
Talks can be viewed at AAAI’s website at https://aaai.org/Conferences/AAAI-20/livestreamed-talks/)