您现在的位置: 实验室首页 >> 最新动态 >> 学术报告 >> 关于美国加州理工学院Dr. Stephan Zheng和 Dr.Qi (Rose) Yu学术报告的通知
关于美国加州理工学院Dr. Stephan Zheng和 Dr.Qi (Rose) Yu学术报告的通知

报告时间:6月29日(周四)9:30 -11:30
报告地点:浙江大学紫金港校区蒙民伟楼402
主持人:蔡登 教授

Title1: Efficient Deep Imitation and Reinforcement Learning in Multi-Agent Environments9:30 -10:30

Abstract: Deep learning and reinforcement learning have been highly successful in training AIagents that perform low-level pattern recognition and short-term decision-making, such as image classification and playing the game of Go. However, many practical problems, such as autonomous driving, intelligent logistics and robotics, require learning behavioral policies that are capable of high-level reasoning in complex environments. In this talk, I will present novel methods to learn policies for 2 challenges in this context: 1) reasoning over long timescales in spatiotemporal multi-agent games and 2) reasoning about cooperation in multi-agent coordination games. In both settings, machine learning faces fundamental scalability and feasibility challenges. To address these, I will present 2 novel deep imitation and reinforcement learning approaches. First, to learn long-term planning in multi-agent games, I will present a class of hierarchical deep learning models that operate on different timescales. To illustrate their effectiveness, I will show that these models are able to learn to move like professional basketball players by imitation from human demonstrations. Second, I will demonstrate “MACE” (Multi-Agent Coordinated Exploration), a technique that improves the sample efficiency of reinforcement learning in games with many agents by jointly learning how to explore and coordinate. Finally, I will discuss ongoing research on improving long-term sequential prediction models via reinforcement learning and nonlinear Lyapunov control.

Bio:Stephan Zheng (www.stephanzheng.com) is a Ph.D. candidate in the Machine Learning Group at Caltech, advised by Professor Yisong Yue. His main research focuses on developing new deep reinforcement learning methods for multi-agent environments. Previously, he has worked on deep imitation learning and robust deep learning. Stephan has published in leading machine learning and computer vision conferences, such as NIPS and CVPR. He was twice a research intern with Google Research and Google Brain. Before machine learning, he worked on theoretical high-energy physics and topological string theory. Stephan obtained his Master’s degrees in Mathematics and Theoretical Physics from the University of Cambridge and Utrecht University, and was a visiting student at Harvard University. He received the 2011 Lorentz prize in Theoretical Physics from the Royal Dutch Academy of Sciences and Arts, and twice the Dutch National Huygens Fellowship.

Title2Scalable structure learning for spatiotemporal analysis10:30 -11:30

Abstract:Spatiotemporal data is ubiquitous in our daily life, including climate,transportation, and social media. Today, data is being collected at an unprecedentedscale. As such, yesterday’s concepts and tools are insufficient to serve tomorrow’sdata-driven decision makers. Particularly, spatiotemporal data often demonstratescomplex dependency structures and is of high dimensionality. This requires newmachine learning algorithms that can handle highly correlated samples, performefficient dimension reduction, and generate structured predictions.In this talk, I will present a set of new machine learning tools for spatiotemporaldata analysis, with an emphasis on structure learning. I will discuss two frameworks.(1) Low-rank tensor learning. Tensors provide natural representations forhigh-ordercorrelations. I will demonstrate scalable tensor methods forspatiotemporal analysis. This includes a geometric rate and linear memorysolver for low-rank tensor learning with theoretical guarantees.(2) Graph convolutional recurrent neural networks. Many spatiotemporal problemsare set in irregular geometry, such as non-grid graphs, which cannot becaptured well by most existing deep learning methods. I will present a noveldeep architecture that can model irregular spatial temporal dependencies. Thismodel achieves significant improvements in traffic forecasting applications.The talk includes joint work with Yan Liu, Cyrus Shahabi, and Yaguang Li.

Bio:Qi (Rose) Yu (www.roseyu.com) is a Ph.D. candidate, Annenberg fellow at theUniversity of Southern California, focusing on machine learning and data analytics.She will join Caltech as a postdoc and then a tenure-track assistant professor inNortheastern University starting 2018 fall. Her research strives to develop machinelearning methods to learn from large-scale spatiotemporal data, specifically in thedomain of computational sustainability and social science. She has over a dozenpublications in leading machine learning/data mining conferences (NIPS, ICML, AAAI,KDD), including best paper nominations. Previously, she was a visiting researcher inStanford University. S he has interned in various industrial research labs includingIBM Watson, Yahoo, Intel Labs and Microsoft. As part of technology transfer, sheco-founded a neighborhood discovery startup and led the team into the USC incubatorprogram. She was n ominated as one of the ``2015 MIT Rising Stars in EECS’’.

[时间:2017-06-20 12:57 点击: 次]
地址:中国·浙江·杭州·余杭塘路866号(310058)
Copyright © 浙江大学CAD&CG国家重点实验室 浙ICP备05074421