20201021-26 图神经网络：深图远算，理胜其辞

报告嘉宾：廖仁杰 (University of Toronto)

报告题目：Deep Learning on Graphs

报告嘉宾：陈思衡 (Mitsubishi Electric Research Laboratories)

报告题目：Bridging graph signal processing and graph neural networks

Panel嘉宾：

唐杰 (清华大学)、石川 (北京邮电大学)、廖仁杰 (University of Toronto)、陈思衡 (Mitsubishi Electric Research Laboratories)、胡玮 (北京大学)

Panel议题：

1. 图神经网络的发展当前面临哪些问题？哪些方面有望在近几年取得突破？

2. 如何基于图神经网络做认知推理？如何看待图神经网络与知识图谱的结合？

3. 如何看待无监督/自监督图神经网络？未来的发展方向及面临的主要困难是？

4. 图神经网络在可解释性方面有什么优势与不足，还能如何增强？

5. 如何在图表征学习中更好的构建与运用图结构？

6. 怎么有效的处理大规模图，怎么有效的处理动态图（time-varying graphs）？

*欢迎大家在下方留言提出主题相关问题，主持人和panel嘉宾会从中选择若干热度高的问题加入panel议题！

报告嘉宾：廖仁杰 (University of Toronto)

报告时间：2020年10月21日(星期三)晚上20:00(北京时间)

报告题目：Deep Learning on Graphs

报告人简介：

Renjie Liao is a PhD student from the machine learning group, University of Toronto. He is jointly supervised by Richard Zemel and Raquel Urtasun. He also works part-time as a Senior Research Scientist in Uber Advanced Technology Group. He is interested in machine learning, computer vision, and self-driving, and has published more than 30 papers at top-tier conferences including NeurIPS/NIPS, ICLR, ICML, CVPR, ICCV, ECCV, ICRA, and CoRL. He recently focused on deep learning on graphs and has co-organized related workshops at ICML 2020, NeurIPS 2019, ICML 2019, and KDD 2019. He will join UBC ECE and CS (affiliated) as an assistant professor in 2021.

个人主页：

http://www.cs.toronto.edu/~rjliao/

报告摘要：

Graphs are ubiquitous in many domains like computer vision, natural language processing, computational chemistry, and computational social science. Although deep learning has achieved tremendous success, effectively handling graphs is still challenging due to their discrete and combinatorial structures. In this talk, I will discuss my recent work which improves deep learning on graphs from both modeling and algorithmic perspectives.

First, I will discuss graph representation learning with a focus on how to effectively capture the multi-scale dependencies in the graph-structured data [1]. I will then describe how to learn a cost function on bipartite graphs using a simple (30 lines of code) differentiable matching algorithm [2]. At last, I will move from discriminative models to generative models and introduce an efficient and scalable deep generative model of graphs [3]. 

参考文献：

[1] Liao, R., Zhao, Z., Urtasun, R. and Zemel, R.S., 2019. Lanczosnet: Multi-scale deep graph convolutional networks. In International Conference on Learning Representations (ICLR).

[2] Zeng, X.*, Liao, R.*, Gu, L., Xiong, Y., Fidler, S. and Urtasun, R., 2019. DMM-Net: Differentiable mask-matching network for video object segmentation. In IEEE International Conference on Computer Vision (ICCV).

[3] Liao, R., Li, Y., Song, Y., Wang, S., Hamilton, W., Duvenaud, D.K., Urtasun, R. and Zemel, R., 2019. Efficient graph generation with graph recurrent attention networks. In Advances in Neural Information Processing Systems (NeurIPS).

报告嘉宾：陈思衡 (Mitsubishi Electric Research Laboratories)

报告时间：2020年10月21日(星期三)晚上20:30(北京时间)

报告题目：Bridging graph signal processing and graph neural networks

报告人简介：

Siheng Chen is a research scientist at Mitsubishi Electric Research Laboratories. Before that, he was an autonomy engineer at Uber Advanced Technologies Group, working on the perception and prediction systems of self-driving cars. Before joining Uber, Dr. Chen was a postdoctoral research associate at Carnegie Mellon University. Dr. Chen received his doctorate in Electrical and Computer Engineering from Carnegie Mellon University in 2016, where he also received two master degrees in Electrical and Computer Engineering (College of Engineering) and Machine Learning (School of Computer Science), respectively. He received his bachelor’s degree in Electronics Engineering in 2011 from Beijing Institute of Technology, China. He is the recipient of the 2018 IEEE Signal Processing Society Young Author Best Paper Award. His co-authored paper received the Best Student Paper Award at 2018 IEEE Global Conference on Signal and Information Processing. His research interests include signal processing, machine learning and autonomous driving.

个人主页：

https://scholar.google.fr/citations?hl=en&user=W_Q33RMAAAAJ

报告摘要：

With the explosive growth of information and communication, data today is generated at an unprecedented rate from various sources, including social, biological and physical infrastructure, among others. Unlike time-series or images, these data possess complex, irregular structures, which can be modeled as graphs. Analyzing graph data requires new concepts and tools to handle the underlying irregular relationships, leading to the emerging fields of graph signal processing and graph neural networks. Graph signal processing (GSP)generalizes the classical signal processing toolbox to the graph domain and provides a series of mathematically-designed operations to process graph signals. On the other hand, graph neural networks (GNN) expand deep learning techniques to the graph domain and provide a data-driven framework to learn from graph signals with graph structures as induced biases. Permeating the benefits of deep learning to the graph domain, graph convolutional networks and variants have attained remarkable empirical success in social network analysis, quantum chemistry and computer vision. Both research direction could potentially complement each other.

In this talk, we will introduce methods that connect these two fields from two perspectives: operations and architectures. From the perspective of operations, we specically discuss the connections between graph sampling theory in GSP and graph pooling in GNN. From the perspective of architectures, we will show the connections between graph scattering transforms and graph convolutional networks. We will also present the related applications to skeleton-based action recognition and multi-agent motion prediction.

Panel嘉宾：唐杰 (清华大学)

嘉宾简介：

唐杰，清华大学计算机系教授、系副主任，获杰青。研究人工智能、认知图谱、数据挖掘、社交网络和机器学习。发表论文300余篇，引用16000余次，获ACM SIGKDD Test-of-Time Award（十年最佳论文）。主持研发了研究者社会网络挖掘系统AMiner，吸引全球220个国家/地区2000多万用户。担任IEEE T. on Big Data、AI OPEN主编以及WWW’21、CIKM’16、WSDM’15的PC Chair。获北京市科技进步一等奖、人工智能学会一等奖、KDD杰出贡献奖。

个人主页：

http://keg.cs.tsinghua.edu.cn/jietang/

Panel嘉宾：石川 (北京邮电大学)

嘉宾简介：

石川，北京邮电大学计算机学院教授、博士研究生导师、智能通信软件与多媒体北京市重点实验室副主任。主要研究方向: 社会网络分析、数据挖掘、机器学习和大数据分析。在IEEE TKDE、ACM TIST、KDD、WWW、AAAI、IJCAI等期刊和会议发表论文100余篇，出版英文专著一部，申请国家发明专利三十余项，研究成果应用到阿里、腾讯、华为等企业。获得ADMA2011/ADMA2018国际会议最佳论文奖，并指导学生获得顶尖国际数据挖掘竞赛IJCAI Contest 2015 全球冠军。研究成果获得省部级奖项2项，获得北京市师德先锋和北京市高等学校青年英才计划支持。

个人主页：

www.shichuan.org

主持人：胡玮 (北京大学)

主持人简介：

Wei Hu is an Assistant Professor at Wangxuan Institute of Computer Technology, Peking University. She received the B.S. degree from University of Science and Technology of China in 2010, and the Ph.D. degree in Electronic and Computer Engineering from the Hong Kong University of Science and Technology in 2015. She was a researcher at Technicolor, France from 2015 to 2017. Her research interests include graph signal processing (GSP), graph neural networks (GNNs) and 3D visual computing. She provided novel inter-disciplinary interpretation and design of GNNs combining with GSP-based mathematical tools. She has regularly published in top image processing / computer vision venues, including TIP, TSP, CVPR, ECCV and ACM MM, with 40+ international journal and conference publications in total. She is the recipient of the Best Student Paper Runner Up Award in ICME 2020, Top 10% Paper Award in ICIP 2014 and MMSP 2013 respectively.

个人主页：

https://www.wict.pku.edu.cn/huwei/

主持人：杨旭 (中科院自动化所)

主持人简介：

杨旭，中国科学院自动化研究所复杂系统管理与控制国家重点实验室副研究员，于2014年获得中国科学院自动化研究所博士学位。研究方向包括图算法、模式分析、计算机/机器人视觉、视觉感-认知发育等。发表国际期刊与会议论文50余篇，包括IEEE TPAMI、IJCV、IEEE T-CYB、IEEE TNNLS, IEEE TSMC-Systems, PR等。申请专利20余项。承担国家自然科学基金委、科技部等部门项目（课题）10余项。担任SCI国际期刊International Journal of Advanced Robotic Systems (IJARS)视觉系统方面编委。带队参加国家自然科学基金委主办的水下机器人目标抓取大赛，于2017年首届获得自主抓取组第1，2018年第二届获得目标识别组第1。获北京市科学技术奖二等奖。

个人主页：

http://people.ucas.edu.cn/~XuYang

20-26期VALSE在线学术报告参与方式：

长按或扫描下方二维码，关注“VALSE”微信公众号 (valse_wechat)，后台回复“26期”，获取直播地址。

特别鸣谢本次Webinar主要组织者：

主办AC：杨旭 (中科院自动化所)、胡玮 (北京大学)

责任AC：严骏驰 (上海交通大学)