Most standard machine models are designed for inputs with a vector feature representation For many structured inputs such as time-series, strings, histograms, and graphs, since there are no explicit features in data, much work has aimed to develop the effective representation of these complex inputs Despite the great success of Deep Learning models have achieved, we aim to develop a generic methodology to learn feature representation directly from "unlabeled " data for structured inputs In the first part of this talk, we will first talk about D2KE, a generic framework to learn a kernel and its embedding from any distance (ICLR 19 submission) In the second part, I will present how to apply D2KE for learning embeddings of multivariate time-series (AIStats 18, oral paper) In the third part, I will discuss how to apply D2KE for learning universal text embedding from pre-trained word embedding (EMNLP 18)

吴凌飞，IBM全球研究院总部（IBM T.J. Watson Research Center) 研究员 ， 威廉玛丽大学计算机系博士，主要研究方向为机器学习，深度学习，表征学习，自然语言处理，大数据。吴博士已经发表20几篇顶尖杂志和会议，包含但不局限于KDD, ICDM, AISTATS, EMNLP, AAAI, ICASSP, SC, SIAM Journal on Scientific Computing, IEEE Transaction on Big Data, and Journal of Computational Physics。吴博士同时也是13项美国专利的发明人。吴博士长期担任其相关研究领域多家国际顶尖杂志的评审，比如SIAM Journal on Scientific Computing, SIAM Journal on Matrix Analysis and Applications, IEEE Transactions on Parallel and Distributed Systems, IEEE Transactions on Knowledge and Data Engineering, IEEE Transactions on Big Data, ACM Transactions on Management Information Systems, Data Mining and Knowledge Discovery, Pattern Recognition, and Future Generation Computer Systems。吴博士同时担任了 IEEE Big Data'18 Tutorial Co-Chair (讲座主席)， 并长期担任AI/ML/DL/DM国际顶会的TPC (技术委员会委员)， 如ICML'19, ICLR'19, WSDM'19, ACML'18, NIPS'18, ICML'18, IJCAI'18, KDD'18, ICDM'18, ICASSP'18, AAAI'18, WSDM'18, ICDM'17, IJCAI'17, BigData'17, KDD'17, AAAI'17, ICSSAP’17, and BigData'16。