The 2018 Conference on Neural Information Processing Systems (NeurIPS) kicked off today in Montréal. In their opening remarks this afternoon the NeurIPS organizing committee announced the conference’s major best paper selections and other awards.
Four submissions shared top honors in the best paper category:
Non-Delusional Q-Learning and Value-Iteration from researchers at Google AI: The paper first identified a fundamental problem in Q-learning, “delusional bias”, and demonstrated its detrimental consequences; then proposed a new policy-consistent backup operator that can fully resolve the problem of delusion. Click the link to read the full paper.
Optimal Algorithms for Non-Smooth Distributed Optimization in Networks from researchers at Huawei Noah’s Ark Lab, Microsoft Research, MSR-INRIA Joint Centre, PSL Research University, and University of Washington. The paper studied the distributed optimization of non-smooth convex functions and proposed two algorithms to solve the problem — multi-step primal-dual (MSPD) and distributed randomized smoothing (DRS). Click the link to read the full paper.
Nearly Tight Sample Complexity Bounds for Learning Mixtures of Gaussians via Sample Compression Schemes from researchers at McMaster University, University of Waterloo, University of British Columbia, and McGill University. The paper proposed a general technique for distribution learning, then employed this technique in the important setting of mixtures of Gaussians. Click the link to read the full paper.
Neural Ordinary Differential Equations from a team of Vector Institute researchers at University of Toronto. The paper parameterized the continuous dynamics of hidden units using an ordinary differential equation (ODE) specified by a neural network and developed a new family of deep neural network models for time-series modeling, supervised learning, and density estimation. Click the link to read the full paper.
The Test of Time Award goes to The Tradeoffs of Large-Scale Learning, a NIPS 2007 paper from researchers at NEC laboratories of America and Google Zurich. The paper developed a theoretical framework that takes into account the effect of approximate optimization on learning algorithms. Click the link to read the full paper.
This year, there were 4,854 papers submitted and sent out for review. The 21 percent acceptance rate is the same as last year. Algorithms surprisingly topped the list of subject areas, followed by deep learning, applications, reinforcement learning & planning, probability methods, theory, optimization, neuroscience & cognitive science, data and etc, and others.
Almost 9000 will attend the week-long conference, which features nine tutorials, 39 workshops, eight competitions, and 20 demos. NeurIPS changed its acronym from “NIPS” last month to avoid a “hostile environment” due to the term being “vulnerable to sexual puns.”
NeurIPS 2018 runs December 2–9 at the Palais des Congrès in Montréal, Canada. Synced will be reporting from the conference throughout the week.
Journalist: Tony Peng, Fangyu Cai | Editor: Michael Sarazen