预告 | 北京大学前沿计算研究中心第二届青年论坛日程来喽！

日程来啦！

08:00-08:30

REGISTRATION

08:30-08:35

Opening Remarks

Baoquan Chen, Executive   Director, CFCS

08:35-09:20

Keynote: 人工智能赋能全球创作与交流平台

马维英，字节跳动公司副总裁

09:20-09:50

Making Virtual Cinematography Smart

Marc Christie, University of Rennes 1, France

09:50-10:15

Group Photo & BREAK

10:15-10:45

Robust Geometry Processing for Computational Fabrication

Qingnan Zhou, Adobe Research

10:45-11:15

Deep Understanding of Urban Mobility from Cityscape Webcams

Shanghang Zhang, Carnegie Mellon University

11:15-11:45

Bayesian Auctions with Efficient Queries

Bo Li, Stony Brook University

11:45-12:15

Matching Algorithms in E-Commerce

Pan Xu, University of Maryland, College Park

12:15-14:00

LUNCH

14:00-14:30

Soft Advantage Learning: A Unified Method that Learns from Demonstrations and Environments

Yang Gao, University of California Berkeley

14:30-15:00

Value Iteration Network

Yi Wu, University of California Berkeley

15:00-15:30

Spatio-Temporal Data Mining and Privacy Protection

Jiaxin Ding, Stony Brook University

15:30-16:00

Recent Advances on Algorithms in Doubling Metrics
Shaofeng Jiang, Weizmann Institute of Science

16:00-16:20

BREAK

16:20-16:50

Deterministic Document Exchange Protocols, and Almost Optimal Binary Codes for Edit Errors

Kuan Cheng, John Hopkins University

16:50-17:20

Neural Relaxation for Reinforcement Learning in Natural Language Processing

Lili Mou, Adeptmind Inc.

17:20-17:50

QANet: Towards Efficient Human-Level Reading Comprehension

Wei Yu, Carnegie Mellon University

17:50

CLOSING

Yizhou Wang, Vice Director, CFCS

论坛报告人看这里！

按日程排序，部分内容有删减

马维英

马维英，现任字节跳动公司副总裁兼人工智能实验室负责人，带领团队在机器学习、计算机视觉、计算机图形学、语音和音乐、自然语言处理、个性化推荐和搜索等领域进行基础研究和核心技术开发。 他的团队所开发的技术通过字节跳动的产品（例如今日头条和抖音）已经在全球范围被数亿日活跃用户使用。他曾在世界级会议和学报上发表过逾300篇论文，并拥有160多项技术专利。他是电气电子工程师学会院士(IEEE Fellow)、美国计算机协会杰出科学家(ACM Distinguished Scientist)及中国“千人计划”专家。他是2008国际互联网大会(WWW)的程序委员会联合主席, 以及2011年国际信息检索大会(SIGIR)的联合主席。2018年7月，马维英入选TOP100的CS计算机科学家，h-index 104，全球排名86，中国排名第2。

人工智能赋能全球创作与交流平台

摘要：字节跳动公司的使命是建立新一代全球信息平台，从内容创作、分发、互动、和交流的每一个环节，用人工智能技术赋能，提升用户体验，促进人类信息与知识交流的效率与深度。例如，通过计算机视觉，自然语言理解和生成技术开发的自动写稿机器人，能够自动理解体育视频并产生新闻播报。通过人工智能辅助内容审核，能够处理每天海量用户生成的各种内容。通过计算机视觉技术在手机端的应用，包括人脸检测和关键点定位、通用物体检测和识别，图像分类、分割、智能化美颜美妆、人体姿态估计、手势识别、手指关节点定位、SLAM等，抖音赋能每个人都能创作出高质量和内容丰富的短视频。在音频内容创作方面，基于深度学习的语音合成系统，应用到新闻播报和小说听书。同时，我们还在积极探索个性化合成技术，包括模拟不同发音人的音色与风格等。音乐是具有高商业价值的内容形式，同时也是构成其它内容的重要元素，因此音乐生成也是我们探索的研究方向。在这个演讲中我将会介绍人工智能在文本、视频、语音、音乐的自动理解和生成技术的最新发展，以及在内容创作和交流上的许多新的应用。

Marc Christie

Marc Christie is an associate professor at University of Rennes 1. His research is focused on virtual cinematography which is the application of real cinematography techniques to virtual 3D environments. The research covers a wide range of challenges like extracting data from real-movies, learning elements of film style (types of transitions, continuity between shots, editing patterns), proposing models and techniques to re-apply the learnt elements to virtual contents, computing camera angles and trajectories as well as optimal edits. He co-authored 40+ conference papers on these topic, and led courses at Eurographics and Siggraph Asia.

Making Virtual Cinematography Smart

Abstract: Cinematography is the transposition of cinematographic principles from real movies, into virtual environments. Yet interactively placing a virtual camera, moving a camera and cutting between different camera angles requires a good knowledge of cinematographic principles, and a good knowledge of the low-level manipulators in 3D modelers.

This talk will present a collection of models, algorithms and techniques to easily encode the cinematographic principles and assist users in the creation of camera shots, camera motions and cuts between shots. We will focus both on fully automated techniques as well as interactive techniques, and will demonstrate results on applications ranging from movie previsualisation (rehearsing a movie before shooting it), to drone cinematography.

Qingnan Zhou

Dr. Qingnan (James) Zhou is a member of the Creative Intelligence Lab in Adobe Research. He received his PhD from Courant Institute of Mathematical Sciences at New York University in 2016 advised by Prof. Denis Zorin. Qingnan's research interests include robust geometry processing, computational fabrication, physics based simulation and cloud computing.  His 3D dataset, Thingi10K, won the 2017 Symposium of Geometry Processing Dataset Award and the 2018 Computer Graphics Forum cover contest.  He is also the main author of the open source geometry processing library PyMesh.

Robust Geometry Processing for Computational Fabrication

Abstract: 3D printing, or addictive fabrication in general, often serves as an important bridge that connects the digital world with the physical world.  Artists often rely on a number of 3D tools to create and analyze their design to ensure that it can withstand the uncompromising test of physics during and after fabrication.  The correctness and robustness of such 3D tools become increasingly crucial for the success of the fabrication.  In this talk, I will address some of the challenges and opportunities related to applying geometry processing algorithms for computational fabrication, including structural analysis, mesh clean up and robust tetrahedral meshing for finite element method.

Shanghang Zhang

Dr. Shanghang Zhang recently received her PhD from Carnegie Mellon University. Her research covers computer vision and deep learning. She especially focuses on domain adaptation, meta learning, and graph convolutional networks. She has been working on traffic video analysis, salient object segmentation, multi-source domain adaptation with adversarial training, and topology adaptive graph convolutional networks, leading to publications on NIPS, CVPR, ICCV, TMM, ICLR, ICIP, ICME, etc. She is the recipient of Adobe Academic Collaboration Fund, Qualcomm Innovation Fellowship (QInF) Finalist Award, and Chiang Chen Overseas Graduate Fellowship. She has been selected to “2018 Rising Stars in EECS”, US. 

Deep Understanding of Urban Mobility from Cityscape Webcams

Abstract: Deep understanding of urban mobility is of great significance for many real-world applications, such as urban traffic management and autonomous driving. Such problem is extremely challenging due to the low spatial and temporal resolution, high occlusion, large perspective, and variable environment conditions of the large-scale videos from the city cameras. In this talk, I will introduce my research works on extracting vehicle counts from such videos, including: 1. a deep multi-task learning framework based on fully convolutional neural networks to jointly learn vehicle density and vehicle count; 2. deep spatio-temporal networks for vehicle counting to incorporate temporal information of the traffic flow; 3. multi-source domain adaptation mechanisms with adversarial learning to adapt the deep counting model to multiple cameras; and 4. meta-learning based lifelong & few-shot vehicle counting. These techniques are organically integrated into the CityScapeEye system that are extensively evaluated and compared to existing techniques on different counting tasks and datasets, with experimental results demonstrating its effectiveness and robustness.

Bo Li

Bo Li is currently a PhD candidate in the Department of Computer Science at Stony Brook University, advised by Jing Chen. His research interests include algorithmic game theory, mechanism design, multi-agent systems, and other problems at the interface between computer science and economics. He received his bachelor's and master's degrees from Ocean University of China under the supervision of Qizhi Fang. His papers were published in several conferences include ICALP, WINE, IJCAI.

Bayesian Auctions with Efficient Queries

Abstract: Generating good revenue is one of the most important problems in Bayesian auction design, and many (approximately) optimal dominant-strategy incentive compatible (DSIC) Bayesian mechanisms have been constructed for various auction settings. However, most existing studies do not consider the complexity for the seller to carry out the mechanism. It is assumed that the seller knows “each single bit” of the distributions and is able to optimize perfectly based on the entire distributions. Unfortunately this is a strong assumption and may not hold in reality: for example, when the value distributions have exponentially large supports or do not have succinct representations. 

In this work we consider, for the first time, the query complexity of Bayesian mechanisms. We only allow the seller to have limited oracle accesses to the players’ value distributions, via quantile queries and value queries. For a large class of auction settings, we prove logarithmic lower-bounds for the query complexity for any DSIC Bayesian mechanism to be of any constant approximation to the optimal revenue. For single-item auctions and multi-item auctions with unit-demand or additive valuation functions, we prove tight upper-bounds via efficient query schemes, without requiring the distributions to be regular or have monotone hazard rate. Thus, in those auction settings the seller needs to access much less than the full distributions in order to achieve approximately optimal revenue.

Pan Xu

Pan Xu is currently a Ph.D. student in the Department of Computer Science at the University of Maryland, College Park. He is very fortunate to be supervised by Dr. Aravind Srinivasan and Dr. John Dickerson. Pan's research interests broadly span the intersection of Algorithms, Operations Research, and Artificial Intelligence. Recently, he focuses on the design of efficient algorithms for offline and online matching models and their applications into various real matching markets, including crowdsourcing marketplaces, ridesharing platforms, and different online recommendation systems. He is the single nominee by the CS Department at UMD for the CMNS Board of Visitors Outstanding Graduate Student Award, 2018.

Matching Algorithms in E-Commerce

Abstract: Matching is a fundamental model in combinatorial optimization. During the last decade, stochastic versions of matching models have seen broad applications in various matching markets emerging in E-Commerce. In this talk, I will first present two basic models, namely offline and online stochastic matching, which are primarily motivated by the online dating and the Internet advertising business respectively. I will first briefly discuss fundamental algorithms for each model. Then I will show several new challenges and our corresponding algorithmic solutions when we can apply these two basic models to different real matching markets, including crowdsourcing marketplaces (Amazon Mechanical Turk), ridesharing platforms (Uber and Lyft), the online food-ordering business (Grubhub) and the Amazon recommendation systems.

Yang Gao

Yang Gao is a 5th year Ph.D. student in the Computer Science Department at UC Berkeley, advised by Professor Trevor Darrell. He is mainly interested in computer vision, robot learning as well as autonomous driving. Before that, Yang graduated from the computer science department at Tsinghua University, where he worked with Prof. Jun Zhu on machine learning problems. He has also interned at Waymo, the self-driving company under Alphabet, Intel with Vladlen Koltun on end to end autonomous driving, as well as in Google Research on natural language processing problems with Dr. Edward Y. Chang and Dr. Fangtao Li.

Soft Advantage Learning: A Unified Method that Learns from Demonstrations and Environments

Abstract: Robust real-world learning should benefit from both demonstrations and interac- tions with the environment. Current approaches to learning from demonstration and reward perform supervised learning on expert demonstration data and use rein- forcement learning to further improve performance based on the reward received from the environment. These tasks have divergent losses which are difficult to jointly optimize, and such methods can be very sensitive to suboptimal demon- strations. We propose a unified reinforcement learning algorithm, Soft Advantage Learning (SAL), that effectively normalizes the Q-function, reducing the Q-values of actions unseen in the demonstration data. SAL learns an initial policy network from demonstrations and refines the policy in the environment, surpassing the demonstrator’s performance. Crucially, both learning from demonstration and interactive refinement use the same objective, unlike prior approaches that combine distinct supervised and reinforcement losses. This makes SAL robust to suboptimal demonstration data, since the method is not forced to mimic all of the examples in the dataset. We show that our unified reinforcement learning algorithm can learn robustly and outperform existing baselines when evaluated on several realistic driving games.

Yi Wu

Yi Wu is now a 5-th year Ph.D. candidate at UC Berkeley advised by Prof. Stuart Russell. He received his B.E. from the special pilot class (Yao class) from Institute of Interdisciplinary Information Sciences, Tsinghua University. Yi's research focuses on how to effectively incorporate human knowledge into AI models to produce both interpretable and generalizable solutions. He is now working on a variety of projects, including deep reinforcement learning, natural language processing and  probabilistic programming.

Value Iteration Network

Abstract: We introduce the value iteration network (VIN): a fully differentiable neural network with a “planning module” embedded within. VINs can learn to plan, and are suitable for predicting outcomes that involve planning-based reasoning, such as policies for reinforcement learning. Key to our approach is a novel differentiable approximation of the value-iteration algorithm, which can be represented as a convolutional neural network, and trained end-to-end using standard backpropagation. We evaluate VIN based policies on discrete and continuous path-planning domains, and on a natural-language based search task. We show that by learning an explicit planning computation, VIN policies generalize better to new, unseen domains.

Jiaxin Ding

Jiaxin Ding is now a Ph.D. candidate in Computer Science, Stony Brook University. His advisor is Professor Jie Gao. He will defend his dissertation in November. Jiaxin Ding received his B.S. in EECS, and B.A. in Economics (double major), Peking University, in 2012. His research interests include spatio-temporal data ming, differential privacy, computational geometry, and Internet of Things. His papers are published in the conferences of IPSN, INFOCOM, MobiHoc, and SIGSPATIAL.

Spatio-Temporal Data Mining and Privacy Protection

Abstract: Huge volume of spatio-temporal data is generated with the advances of sensing techniques and computing capability. It provides us great opportunities to study mobility patterns to enhance our living qualities while it also raises privacy issues when collecting individuals' trajectories. In this presentation, I describe geometric algorithms to sense trajectories, mine mobility patterns, and protect individuals' privacy. 

Instead of using location devices equipped by mobile entities (vehicles, pedestrians, etc.), we employ checkpoints (roadside units, WiFi access points, cellular towers, etc.) to record appearances of the mobile entities. We apply a  hierarchical data structure based on a succinct differential private MinHash signature to frequent traffic patterns efficiently.  

We also demonstrate an adversary can re-identify individuals through frequent location attack, co-location attacks and motif attacks. We introduce an approach mixing ID during the co-location events. The performance of the approach is effective theoretically and empirically to defeat statistical attacks.

Shaofeng Jiang

Shaofeng Jiang is a postdoctoral research fellow at the Foundations of Computer Science Group, Weizmann Institute of Science, hosted by Prof. Robert Krauthgamer. Before joining Weizmann, he obtained the PhD from the University of Hong Kong at 2017, and Bachelor’s degree from Shandong University at 2013. He is working on theoretical computer science, currently with an emphasis on algorithms for big data, algorithms for bounded dimensional metric spaces and online algorithms. Many of his research has been published in top venues such as SODA and FOCS. His thesis was nominated for the best thesis award in HKU, and he won the nomination award of MSRA Fellowship 2015.

Recent Advances on Algorithms in Doubling Metrics

Abstract: Doubling dimension measures the intrinsic dimensionality of a metric space. The most well known doubling metrics are the bounded dimensional Euclidean spaces. Moreover, many other important distance measures that are widely used in fields such as computer vision, machine learning and data analysis, are also doubling. Algorithms for doubling metrics work for wide ranges of metric spaces, and can often be more efficient than algorithms that only use the ambient dimension.

In this talk, we give an introduction to the notion of doubling dimension, and we discuss several recent advances on algorithms in doubling metrics. In particular, we introduce a series of my works on approximation algorithms in doubling metrics, and also a very recent work on data reduction for clustering problems. We conclude with future directions.

Kuan Cheng

Kuan Cheng is a PhD candidate at Johns Hopkins University, Computer Science Department, advised by Prof. Xin Li, working on Randomness and Combinatorics in Computation, and their applications in Complexity Theory, Information Theory, etc., also interested in Machine Learning, Networks and other topics in Computer Science. 

Deterministic Document Exchange Protocols, and Almost Optimal Binary Codes for Edit Errors

Abstract: We study two basic problems regarding edit error, i.e. document exchange and error correcting codes for edit errors (insdel codes). For message length n and edit error upper bound k, it is known that in both problems the optimal sketch size or the optimal number of redundant bits is Θ(k log n/k). However, known constructions are far from achieving these bounds.

We significantly improve previous results on both problems. In obtaining our results we introduce the notion of ε-self matching hash functions and ε-synchronization hash functions. We believe our techniques can have further applications in the literature.

Lili Mou

Lili Mou is currently a research scientist at AdeptMind Research. Lili received his BS and PhD degrees in 2012 and 2017, respectively, from School of EECS, Peking University. Then, he worked as a postdoctoral fellow at the University of Waterloo, Canada. His research interests include deep learning applied to natural language processing as well as programming language processing. He has publications at top conferences and journals like AAAI, ACL, CIKM, COLING, EMNLP, ICML, IJCAI, INTERSPEECH, and TACL (in alphabetic order).

Neural Relaxation for Reinforcement Learning in Natural Language Processing

Abstract: Reinforcement learning (RL) is a widely applied algorithm in natural language processing (NLP), as almost all language components are discrete (e.g., characters, words, and sentences). However, it is known that an RL system is difficult to train and sensitive to initialization. In this talk, I will present a framework of neural relaxation for reinforcement learning. Our intuition is that a fully neuralized system is differentiable, and thus end-to-end learnable. If we design a partially interpretable neural network and transfer its knowledge to a discrete system, it can serve as a meaningful initial policy for RL. We accomplish such knowledge transfer by step-by-step imitation learning, and then, the discrete system improves its policy by RL in a trial-and-error fashion. We apply the framework into two applications: unsupervised syntactic parsing and semantic parsing.

Experimental results show that RL alone is difficult to get rid of the initial policy, whereas the neural relaxed model lacks full interpretation and has low performance. Our proposed framework is able to recover latent parsing actions with high accuracy in both applications. 

Wei Yu

(Adams) Wei Yu a PhD candidate in Machine Learning Department of SCS at CMU, advised by Jaime Carbonell and Alex Smola. His research interest is in artificial intelligence, encompassing deep learning, large scale optimization and natural language processing. The main theme of his research is to accelerate AI by designing efficient models and algorithms. His research will be/is/was supported by Snap PhD Fellowship, NVIDIA PhD fellowship, CMU Presidential Fellowship and Siebel Scholarship.

QANet: Towards Efficient Human-Level Reading Comprehension

Abstract: In this talk, I will introduce a neural machine reading comprehension system, called QANet. Unlike other frameworks, it discards recurrent neural networks, but only uses convolution and self-attention as the encoder, and is by far the deepest neural network in the NLP domain. As a result, it is substantially faster than all the existing models. Besides, combined with a novel data augmentation approach via cyclic-translation, QANet also achieves No.1 performance in the competitive Stanford Question and Answer Dataset (SQuAD), as of Aug 2018. Notably, it outperforms human on the exact match metric by a large margin.

听众报名请扫下方小程序码
（名额有限，报满为止）

想加入我们？请戳下方链接或点击“阅读原文”！