More about HKUST
HKUST CSE Special Seminar
Title: "HKUST CSE Special Seminar" Date: 14 April (Friday), 2023 Time: 10:00 - 13:00 (HKT) Venue: Zoom (https://hkust.zoom.us/j/94531153482?pwd=UUU4QlpLT3p0bmdSKzVoYlhKWFd6Zz09) Meeting ID: 945 3115 3482 Passcode: 328985 Host: Dr. Amir GOHARSHADY and Dr. Xiaojuan MA ============================================= *** Speaker 1 *** Dr. Huan Zhao Senior Researcher of 4Paradigm Inc. Time: 10:00 -11:00 Topic: Advances in Automated Graph Neural Networks: Research and Applications from an AI Startup Q&A: 11:00-11:30 *** Speaker 2 *** Prof. Umang Mathur Assistant Professor School of Computing at the National University of Singapore Time: 11:30-12:30 Topic: Rethinking Timestamping in Distributed Computing Through a Data-Structure Lens Q&A: 12:30-13:00 ************* Dr. Huan Zhao Topic: Advances in Automated Graph Neural Networks: Research and Applications from an AI Startup Abstract: In recent years, graph neural networks (GNNs), e.g., GCN, GAT, and JK-Network, have been the state-of-the-art (SOTA) methods for various tasks in GSD, e.g., product recommendation, and molecular property prediction. However, due to diverse graphs in real-world applications, currently, no human-designed GNNs can consistently perform well, and thus a huge amount of human expertise and computational resources have to be invested to obtain well-performing GNN architectures. In this talk, I will introduce our exploration of automated graph neural networks (AutoGraph). Specifically, we use the recently popular neural architecture search (NAS) for graph architecture design. Specifically, two works based on differentiable neural architecture search algorithms will be introduced, which can search for better GNN architectures on different datasets including node-level and graph-level tasks, and show that they can effectively and efficiently obtain well-performing data-specific GNN architectures. Moreover, the proposed method for graph classification has achieved 1st place in the well-known open graph benchmark (OGB). These two works have been published in top-tier data mining conferences, i.e., CIKM 2021 and WebConf 2022. Finally, a real-world application of AutoGraph to biomedicine scenario, i.e., predicting Double-strand DNA breaks, will be briefly introduced. Biography: Dr. Huan Zhao is a senior researcher in 4Paradigm Inc., an AI startup in China, leading the research on automated graph neural networks (AutoGraph) in the company. Prior to 4Paradigm, He worked as a research intern and senior algorithm engineer at Alibaba from November 2017 to July 2019. He obtained his Doctor Degree at the Department of Computer Science and Engineering, HKUST in Jan. 2019. His research interests include recommender system, graph representation learning, and automated machine learning. He has published more than 30 top-tier conference and journal papers, including KDD, WebConf, ACL, and AAAI etc. Besides research, he also leads a team in 4Paradigm to deliver AutoGraph algorithms to real-world applications, e.g., retailing recommendation, financial fraud detection, bioinformatics, intelligent manufacturing, etc., from the business partners of 4Paradigm. ****************** Prof. Umang Mathur Topic: Rethinking Timestamping in Distributed Computing Through a Data-Structure Lens Abstract: Many applications in distributed and concurrent computing crucially rely on logical timestamping. This involves assigning timestamps to different events in the execution of a system so that the ordering induced on the assigned timestamps is isomorphic to the underlying causal ordering on events in the execution. In other words, the timestamps of two events can be used to infer the causal ordering between them. In this talk I will first argue that timestamping (i.e., the process of computing and assigning timestamps to all events) should be viewed from a data structure perspective. This perspective allows us to ask questions like "what is the most optimal data structure for timestamping"? The most popular data structure is the vector clock data structure, which stores an integer for each node/thread in a flat array. The two basic updates on vector clocks, namely join and copy, perform exactly k integer operations, where k is the number of threads/processes. In applications where timestamps need to be computed at every event, such vector clock updates can be a computational bottleneck, especially when k is large. Next, I will describe tree clocks, a new data structure that replaces vector clocks for computing causal orderings in program executions. Joining and copying tree clocks takes time that is roughly proportional to the number of entries being modified, and hence the two operations do not suffer the a-priori I?(k) cost per application. I will then discuss an optimality result in the context of an application of my interest, namely data race detection. In shared memory multi-threaded programs synchronising over locks, the happens-before (HB) partial order on events in an execution is a classic partial order used to infer the presence of data races dynamically. For computing the HB partial order, tree clocks are optimal, in the sense that no other data structure can lead to smaller asymptotic running time. Finally, for the practically inclined, I will present empirical evidence that this data structure can significantly speed up applications such as data race detection, and for the theoretically inclined, I will discuss some future extensions and applications. Biography: Umang Mathur is an Assistant Professor at the National University of Singapore. He received his PhD from the University of Illinois at Urbana Champaign and was an NTT Research Fellow at the Simons Institute for the Theory of Computing at Berkeley. His research broadly centres on developing techniques inspired from formal methods and logic for answering design, analysis and implementation questions in programming languages, software engineering and systems. He has recieved a Google PhD Fellowship, an ACM SIGSOFT Distinguished Paper Award at ESEC/FSE'18. Best Paper Award at ASPLOS'22 and an ACM SIGPLAN Distirnguished Paper Award at POPL'23 for his work on designing techniques and tools for analyzing concurrent software. More details can be found at: https://www.comp.nus.edu.sg/~umathur/