BEGIN:VCALENDAR VERSION:2.0 PRODID:-//https://techplay.jp//JP CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALDESC:Talks by Hanjun Dai (Google Brain) and Feng Liu (Australian Ar tificial Intelligence Institute\, UTS) X-WR-CALNAME:Talks by Hanjun Dai (Google Brain) and Feng Liu (Australian Ar tificial Intelligence Institute\, UTS) X-WR-TIMEZONE:Asia/Tokyo BEGIN:VTIMEZONE TZID:Asia/Tokyo BEGIN:STANDARD DTSTART:19700101T000000 TZOFFSETFROM:+0900 TZOFFSETTO:+0900 TZNAME:JST END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:809263@techplay.jp SUMMARY:Talks by Hanjun Dai (Google Brain) and Feng Liu (Australian Artific ial Intelligence Institute\, UTS) DTSTART;TZID=Asia/Tokyo:20210309T100000 DTEND;TZID=Asia/Tokyo:20210309T120000 DTSTAMP:20260519T094033Z CREATED:20210222T140030Z DESCRIPTION:イベント詳細はこちら\nhttps://techplay.jp/event/80926 3?utm_medium=referral&utm_source=ics&utm_campaign=ics\n\nTalk1: Improved Generative Modeling of Structure Data\nSpeaker: Dr. Hanjun Dai (Google Br ain)\n\nAbstract:\nGenerative modeling remains challenging for discrete s tructured data like program trees or\nmolecule graphs. The most commonly used model for such data is the autoregressive one\, thanks\nto its tract ability. However\, in some situations it may suffer from scalability and expressiveness\nissues\, due to its sequential nature and parameter shari ng in deep models.\nIn this talk\, we will share our recent works on addr essing these two issues that come from\nautoregressive modeling. In the f irst part [1]\, we introduce a scalable autoregressive model for\ngenerat ing graph structures\, where it reduces the training synchronizations fro m O(n) to O(log n)\,\nand inference cost from O(n^2) to O(n*log n). In th e second part [2]\, we propose a local search\nmodel with latent variable s that extends the autoregressive model in the context of learning\nenerg y based models for discrete structured data. We demonstrate the effective ness of current\nworks with real-world applications\, including data augm entation\, program synthesis and software\ntesting.\n\nReferences:\n[1] S calable Deep Generative Modeling for Sparse Graphs\, Dai et.al\, ICML 202 0\n[2] Learning Discrete Energy-based Models via Auxiliary-variable Local Exploration\, Dai et.al\,\nNeurIPS 2020\n\nBio:\nHanjun Dai is currently a research scientist at Google Research\, Brain Team. He obtained his Ph D\nfrom Georgia Institute of Technology\, advised by Prof. Le Song. His r esearch focuses on deep\nlearning with structured data\, combinatorial op timization\, generative modeling\, and the\napplications in chemistry\, b ioinformatics\, programming and natural languages. During his PhD he\nhas also extended his research work through internships with Amazon AI\, Ope nAI and DeepMind.\nHe has published over 30 papers in top-tier conference s and journals\, while his work has been\nrecognized by AISTATS 2016 best student paper\, best paper in Recsys Workshop on Deep Learning\nfor Reco mmender System 2016 and best paper in NIPS 2017 Workshop on Machine Learn ing for\nMolecules and Materials.\n\nTalk2: Towards Trustworthy Transfer Learning: Learning from the Wild\nSpeaker: Dr. Feng Liu (Australian Artif icial Intelligence Institute\, UTS)\n\nAbstract:\nTransfer learning aims to leverage knowledge from domains with abundant labels (i.e.\, source\nd omains) to help train a good classifier or clustering model for the domai ns with insufficient/no\nlabels (i.e.\, target domains). Although recent research on transfer learning has shown its ability to\ntransfer knowledg e from a source domain to a target domain\, most works require some unrea listic\nassumptions to ensure their efficacy. Namely\, existing transfer learning methods still face several\nunsolved and challenging problems in the real world.\nIn this talk\, I will first present three orthogonal di rections of trustworthy transfer learning\,\nincluding 1) the necessity o f transfer learning\, 2) transfer learning under the imperfection of\nsou rce domains\, 3) transfer learning under the imperfection of target domai ns. Then\, I will\nintroduce recent advances in the three directions. Fin ally\, promising future works are presented\ntowards the trustworthy tran sfer learning.\n\nBio:\nDr. Feng Liu is a machine learning researcher wit h research interests in transfer learning and\nhypothesis testing. His lo ng-term goal is to develop intelligent systems that can learn knowledge\n from massive related but different domains automatically.\nCurrently\, he is a postdoctoral researcher at the Australian Artificial Intelligence I nstitute (AAII)\,\nUniversity of Technology Sydney (UTS)\, Australia\, an d the recipient of Australian Laureate\npostdoctoral fellowship. He recei ved his Ph.D. degree in computer science at UTS-AAII in 2020\,\nadvised b y Dist. Prof. Jie Lu and Prof. Guangquan Zhang.\nHe was a research intern with the AI Residency Program at RIKEN Center for Advanced Intelligence\ nProject (RIKEN-AIP)\, working on the trustworthy domain adaptation proje ct with Prof. Masashi\nSugiyama\, Dr. Gang Niu\, and Dr. Bo Han. He visit ed Gatsby Computational Neuroscience Unit at\nUCL and worked on the hypot hesis testing project with Prof. Arthur Gretton\, Dr. Danica J.\nSutherla nd and Wenkai Xu.\nHe has served as program committee (PC) members for Ne urIPS\, ICML\, ICLR\, AISTATS\, ACML. He\nalso serves as a reviewer for m any academic journals\, such as IEEE-TPAMI\, IEEE-TNNLS\, IEEE-TFS\,\nand AMM. He has received the AAII best student paper award (2020)\, UTS-FEIT HDR Research\nExcellence Award (2019)\, Best Student Paper Award of FUZZ -IEEE (2019)\, and UTS Research\nPublication Award (2018). LOCATION:オンライン URL:https://techplay.jp/event/809263?utm_medium=referral&utm_source=ics&utm _campaign=ics END:VEVENT END:VCALENDAR