AI academics from around the world are pouring into New Orleans for the Thirty-Second AAAI Conference on Artificial Intelligence. The five-day event is an all-inclusive technical overview of AI research frontiers. Program chairs are Sheila McIlraith from the University of Toronto and Cornell ‘s Kilian Weinberger.
Since its inception in 1979 as the “American Association of Artificial Intelligence,” the annual gathering has expanded scope to encourage international participation. In 2012 the conference was first held outside the US, in Toronto, Canada; and last year organizers shifted the schedule to avoid conflicting with the Chinese New Year and better accommodate Chinese academics.
Compared to other prestigious conferences like the IJCAI, ICML, and NIPS, the AAAI takes a broader approach with its topic selection, which includes search, planning, knowledge representation, reasoning, NLP, robotics and perception, multiagent systems, statistical learning, and deep learning.
2018 Breaks Records
The AAAI 2018 selection committee received a record high of over 3,800 paper submissions, accepting 933 for an acceptance rate of just under 25%, down about 5% from last year when 789 papers made it out of 2590. The number of AAAI paper submissions has risen steadily since 2013.
A word-frequency search of the papers reveals this year’s top keywords: (in descending order) learning, multi, neural, deep, networks/network, adversarial, attention, model, data, detection.
Meanwhile, keywords “data” and “information” have decreased in use frequency compared to last year.
AAAI 2017 registered a high of 1,692 participants, and we are bound to see an increase this year.
Organizers identified Human-Computer Interaction (HCI) as an Emerging Topic this year. There are four talks and 21 technical papers on HCI, focused on trustable and explainable AI, teamwork/team formation, human-aware planning and behaviour prediction, planning and decision support for human-machine teams, human-agent negotiation, human-robot/negotiation, human-robot/agent interaction, human-in-the-loop/learning, human computation, and human and AI communication protocols.
The AAAI’s first Oxford-style debate on “Advances in Machine Learning have displaced the need for logic in AI” is expected to ignite discussion in the Twitterverse. Debating “for” will be former AAAI chair Tom Dietterich and Bart Selman from Cornell, while Gary Marcus from NYU and IBM researcher Francesca Rossi will argue “against.”
Earlier this month, Dietterich responded to Marcus’ Critique of Deep Learning with a volley of contentious Tweets. Their first public debate will certainly be fiery.
AAAI 2018 will also present the first AAAI/ACM Conference on AI, Ethics, and Society.
University of Alberta and Oxford Win Outstanding Paper Awards
The AAAI-18 Outstanding Paper Award was won by the University of Alberta’s Memory-Augmented Monte Carlo Tree Search, which proposes a new method called Memory-Augmented Monte Carlo Tree Search (M-MCTS), combining the original MCTS algorithm with a memory framework, to provide a memory-based online value approximation. Performance was evaluated in the game of go, showing better results than vanilla MCTS.
Paper co-author Jincheng Mei says their next application scenario for MCTS “is complex reinforcement learning tasks, for example in strategy games.” Another of the co-authors, Dr. Martin Müller, told Synced, “we are also applying MCTs in the game of Hex, with PhD student Chao Gao and Dr. Ryan Hayward.” It’s noteworthy that Dr. Müller mentored David Siler and Aja Huang, the first and second authors of AlphaGo’s celebrated Nature paper.
The Outstanding Student Paper award went to Oxford University’s Counterfactual Multi-Agent Policy Gradients, which features a new reinforcement learning method that can efficiently learn decentralized policies in cooperative multi-agent systems. The multi-agent actor-critic method called Counterfactual Multi-Agent (COMA) was evaluated in StarCraft games, which is a challenging reinforcement learning benchmark task.
Percy Liang – How Should We Evaluate Machine Learning for AI?
AAAI Submission Groups
AI and the Web (AIW)
Cognitive Modeling (CM)
Cognitive Systems (CS)
Computational Sustainability and AI (CSAI)
Game Theory and Economic Paradigms (GTEP)
Game Playing and Interactive Entertainment (GPIE)
Heuristic Search and Optimization (HSO)
Human-AI Collaboration (HAC)
Human-Computation and Crowd Sourcing (HCC)
Humans and AI (HAI)
Knowledge Representation and Reasoning (KRR)
Machine Learning Applications (MLA)
Machine Learning Methods (ML)
Multiagent Systems (MAS)
NLP and Knowledge Representation (NLPKR)
NLP and Machine Learning (NLPML)
NLP and Text Mining (NLPTM)
Planning and Scheduling (PS)
Reasoning under Uncertainty (RU)
Search and Constraint Satisfaction (SCS)
Journalist: Meghan Han | Editor: Michael Sarazen