7 papers won the Outstanding Paper Award, Tsinghua, Renmin University, Zhejiang University and other domestic universities on the list, Academician Zhang Cymbals, Professor Zhu Jun and other collaborative papers were selected!
The author | Ailleurs
Editor| Chen Caixian
Today, ICLR2022 announced the results of the Outstanding Paper Award on the official website, a total of 7 papers were awarded, the domestic universities include Tsinghua University, Chinese University, Zhejiang University, Chongqing University, foreign universities and institutions on the list are Google Research, Antwerp University, Stanford University, Cornell University, University of Toronto, DeepMind, etc.
This year, the ICLR received 3391 submissions, receiving 1095, with an acceptance rate of 32.3%, including 54 papers received as Oral, 176 papers received as Spotlight, and 7 papers received the Outstanding Paper Award for their excellent organization, insight, creativity and potential lasting impact. Among them, there are 3 Papers by Chinese People, and the papers cooperated by Academician Zhang Cymbals, Professor Zhu Jun and others have won the Outstanding Paper Award. In addition, 3 papers were nominated for outstanding papers.
1
A Chinese-winning paper
The award-winning paper by Academician Zhang Cymbals, Professor Zhu Jun and others is "Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models".
Address of the paper: https://openreview.net/pdf?id=0xiJLKH-ufZ
Reasons for this work:
The Defusion probabilistic model (DPM) is a powerful class of generative models that is a rapidly evolving research topic in the field of machine learning. This article aims to address the inherent limitation of the DPM model, which is that the optimal inverse variance in the DPM model is calculated slowly and costly. The authors first show a surprising result that both the optimal inverse variance of DPM and the corresponding optimal KL divergence have their parsed form of a score function. They then proposed a novel and elegant framework for training-free inference: Analytic-DPM, a form of analysis that uses Monte Carlo methods and pre-trained score-based models to estimate variance and KL divergence.
This paper has significant implications in terms of both theoretical contributions (showing that the optimal inverse variance and KL divergence of DPM have analytical forms) and practical benefits (proposing training-free inferences for various DPM models) and is likely to influence future DPM studies.
Picture note: Zhang Cymbals
Zhang Cymbal, Professor of the Department of Computer Science of Tsinghua University, Academician of the Chinese Academy of Sciences, Fellow of CCF, Winner of the 2014 CCF Lifetime Achievement Award, and one of the founders of the field of Chinese Engineering Intelligence. He was the Deputy Director of the Academic Degree Committee and is currently a Technical Advisor to Microsoft Research Asia.
Academician Zhang is engaged in theoretical research such as artificial intelligence, artificial neural networks, and machine learning, and these theories are applied to the research of pattern recognition, knowledge engineering and robotics. In these areas, he has published more than 200 academic papers and 5 (or chapters) monographs (in both English and Chinese). His monograph won the Special Prize for Outstanding Academic Monographs issued by the Publishing House of Colleges and Universities of the State Education Commission. His scientific research achievements have won the ICL European Artificial Intelligence Award, the third prize of the National Natural Science Award, the third prize of the National Science and Technology Progress Award, the first and second prizes of the State Education Commission's Scientific and Technological Progress Award, the first prize of the Ministry of Electronics Industry's Scientific and Technological Progress Award, and the first prize of the Science, Technology and Industry Commission for National Defense. In addition, he participated in the creation of the State Key Laboratory of Intelligent Technologies and Systems, where he served as director from 1990 to 1996. From 1987 to 1994, he served as an expert of the national "863" high-tech program intelligent robot subject matter expert group.
In the past 30 years, he has proposed the quotient space theory of problem solving, and proposed the method of mutual transformation, synthesis and reasoning between multi-granular spaces on the basis of the mathematical model of quotient space. Computational complexity analysis for hierarchical solutions to problems and methods to reduce complexity are proposed. The theory and the corresponding new algorithms have been applied to different fields, such as statistical heuristic search, topological dimensionality reduction method of path planning, time programming based on relationship matrix, and multi-granular information fusion, etc., all of which can significantly reduce the computational complexity. The theory has now become one of the main branches of granular computing. On artificial neural networks, he proposed learning algorithms based on planning and point set coverage. These top-down structural learning methods have significant advantages over traditional bottom-up search methods in many ways.
Academician Zhang Cymbal has been teaching at Tsinghua For more than 60 years, he is already full of peach and plum, and has become an absolute authority in his research field. Although he is over the age of Hua Jia, he is still in good spirits, and now he still insists on writing papers as the first author, which is admirable.
Photo note: Zhu Jun
Zhu Jun is a professor in the Department of Computer Science of Tsinghua University, the director of the Basic Theory Research Center of the Institute of Artificial Intelligence of Tsinghua University, and the chief scientist of machine learning of the Beijing Zhiyuan Artificial Intelligence Research Institute. He won the Tencent Science Exploration Award, was selected as the MIT TR35 China Pioneer, the second IEEE AI 10 to Watch scholar in Asia, the leader of the national "10,000 Talents Program", the first deputy editor of PAMI in China, and the deputy editor and editorial board member of IEEE TPAMI. He has served as the regional co-chair of ICML2014, and the chairman of ICML, NIPS, IJCAI, AAAI and other fields more than 20 times.
Professor Zhu Jun's research work revolves around the basic theory, efficient algorithm and application of machine learning, focusing on the combination of theory and practical problems. He has published more than 100 papers in top international conferences and journals on machine learning for many years, such as ICML, NIPS, IJCAI, AAAI, JMLR, PAMI, etc. The research work has been supported by the National 973 Program, the Natural Science Foundation of China And youth fund and key funds, and has been selected into the "Tsinghua University 221 Basic Research Talent Support Program".
Another winning Paper by Chinese came from Stanford University: Comparing Distributions by Measuring Differences that Affect Decision Making.
Address of the paper: https://openreview.net/forum?id=KB5onONJIAU
Why this work was awarded: This paper proposes a new class of differences that can compare two probability distributions for decision-making tasks based on optimal losses. The authors demonstrate that the proposed method has better testing capabilities than competitive benchmarks on various benchmarks. The Commission believes that the method is not only cleverly thought out and has special empirical significance, it allows users to directly specify their preferences when comparing distributions by deciding on losses, which means that the level of interpretability will be improved.
Shengjia Zhao is a PhD student in the Department of Computer Science at Stanford University, whose research interests include predictive models and autonomous agents, probabilistic deep learning, and uncertainty quantification. His collaborative paper with Xu Yilun, a 2016 Turing class undergraduate student at Peking University, and others has been accepted by iclr2020 "full score".
Photo caption: Shengjia Zhao
His mentor is Stefano Ermon, who won the IJCAI2018 "Computer and Thought Award" and has made an important impact on his research in probabilistic reasoning, machine learning and decision-making. Another of his gifted students, Song Biao, won the ICLR 2021 Outstanding Paper Award for his paper, and Song Biao was sent to Tsinghua University at the age of 14, majoring in mathematical physics as an undergraduate, and studied with Zhu Jun and others before going to Stanford University for a doctorate in 2016.
The third winning Paper by Chinese is "Neural Collapse Under MSE Loss: Proximity to and Dynamics on the Central Path," by Xiaoyan Han, a PhD student at Cornell University.
Address of the paper: https://arxiv.org/abs/2106.02073
Reasons for this work:
This paper puts forward new theoretical insights into the phenomenon of "neural collapse" in the current deep network training paradigm. This paper demonstrates a new mean squared error (MSE) loss decomposition method to analyze each component of the loss under a neural collapse, rather than the mathematically more difficult to analyze cross-entropy loss. By studying the reforming gradient flow along the central path, the authors derived precise kinetic methods for predicting neural collapse. This paper provides novel and enlightening theoretical insights for empirical training in understanding deep networks.
Photo: Xiaoyan Han
2
A Chinese author of an honorary nomination paper
PiCO: Contrastive Label Disambiguation for Partial Label Learning, one of the papers nominated for outstanding paper honors, is Haobo Wang, a doctoral student from the School of Computer Science and Technology of Zhejiang University.
Address of the paper: https://arxiv.org/abs/2201.08984
Reasons for this work:
This paper examines partial label learning (PLL) with the aim of reducing the performance gap between PLLs and supervised peers by addressing two key challenges of PLL representation learning and label disambiguation in a coherent framework. The authors propose a new framework, PiCO, that combines contrast learning and prototyping-based label disambiguation. This article gives interesting theoretical explanations to prove its framework from the perspective of Expectation Maximization (EM). The empirical results are particularly impressive, as PiCO is significantly superior to current state-of-the-art technology in PLL, even achieving results comparable to fully supervised learning.
Attached: Full list of papers nominated for outstanding paper awards and honors View address
https://blog.iclr.cc/2022/04/20/announcing-the-iclr-2022-outstanding-paper-award-recipients/
Leifeng Network