BEGIN:VCALENDAR VERSION:2.0 PRODID:-//https://techplay.jp//JP CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALDESC:[23rd AIP Open Seminar] Talks by Nonconvex Learning Theory Tea m X-WR-CALNAME:[23rd AIP Open Seminar] Talks by Nonconvex Learning Theory Tea m 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:812787@techplay.jp SUMMARY:[23rd AIP Open Seminar] Talks by Nonconvex Learning Theory Team DTSTART;TZID=Asia/Tokyo:20210428T150000 DTEND;TZID=Asia/Tokyo:20210428T170000 DTSTAMP:20260421T063011Z CREATED:20210326T060019Z DESCRIPTION:イベント詳細はこちら\nhttps://techplay.jp/event/81278 7?utm_medium=referral&utm_source=ics&utm_campaign=ics\n\nNonconvex Learni ng Theory Team (https://aip.riken.jp/labs/generic_tech/nonconvex_learn_th eory/?lang=en) at RIKEN AIP\n\nSpeaker 1: Takafumi Kanamori (25 min)\nTit le: Overview and Recent Development in the research activity of Non-Conv ex Learning Theory Team\nAbstract: The main target of our team is to deve lop learning algorithms to deal with complex data and to establish theore tical foundations of statistical learning. In this talk\, we will introdu ce recent development including statistical inference of probability dist ribution on complex domains\, kernel-based feature extraction method for complex data\, the fundamental theory of transfer learning\, and so on. \ n\nSpeaker 2: Kosaku Takanashi (25 min)\nTitle: Nonlinear Ensemble Method s for Time Series Data\nAbstract: We propose a class of ensemble methods that nonlinearly synthesizes multiple sources of information– such as p redictive distributions– in a sequential\, time series context. To unde rstand its finite sample properties\, we develop a theoretical strategy b ased on stochastic processes\, where the ensembled processes are expresse d as stochastic differential equations\, evaluated using Itô’s lemma. We determine the conditions and mechanism for which this class of nonline ar synthesis outperforms linear ensemble methods. Further\, we identify a specific form of nonlinear synthesis that produces exact minimax predict ive distributions for Kullback-Leibler risk and\, under certain condition s\, quadratic risk. A finite sample simulation study is presented to illu strate our results. This is a joint work with Kenichiro McAlinn from Temp le university．\n\nSpeaker 3: Yuichiro Wada (25 min)\nTitle: Spectral Em bedded Deep Clustering\nAbstract: We propose a clustering method based on a deep neural network. Given an unlabeled dataset and the number of clus ters\, our method groups the dataset into the given number of clusters in the original space. We use a conditional discrete probability distributi on defined by a deep neural network as a statistical model. Our strategy is first to estimate the cluster labels of unlabeled data points selected from a high-density region. Then secondly\, by using the estimated label s and the remaining unlabeled data points\, we conduct semi-supervised le arning to train the model. Lastly\, by using the trained model\, we estim ate cluster labels of the remaining unlabeled data points. We conduct num erical experiments on five commonly used datasets to confirm the effectiv eness of the proposed method. This talk is based on a paper Entropy 2019\ , 21(8)\, 795\, with Shugo Miyamoto\, Takumi Nakagawa\, Leo Andeol\, Wata ru Kumagai and Takafumi Kanamori.\n\nSpeaker 4: Hironori Fujisawa (25 min )\nTitle: Transfer learning via L1 regularization\nAbstract: Machine lear ning algorithms typically require abundant data under a stationary enviro nment. However\, environments are nonstationary in many real-world applic ations. Critical issues lie in how to effectively adapt models under an e ver-changing environment. We propose a method for transferring knowledge from a source domain to a target domain via L1 regularization in high dim ension. We incorporate L1 regularization of differences between source pa rameters and target parameters\, in addition to an ordinary L1 regulariza tion. Hence\, our method yields sparsity for both the estimates themselve s and changes of the estimates. The proposed method has a tight estimatio n error bound under a stationary environment\, and the estimate remains u nchanged from the source estimate under small residuals. Moreover\, the e stimate is consistent with the underlying function\, even when the source estimate is mistaken due to nonstationarity. Empirical results demonstra te that the proposed method effectively balances stability and plasticity . This is a joint work with Masaaki Takada (Toshiba Cooperation). This ta lk is based on a paper accepted at NeurIPS 2020.\n\n\n\nAll participants are required to agree with the AIP Open Seminar Series Code of Conduct.\n Please see the URL below.\nhttps://aip.riken.jp/event-list/termsofpartici pation/?lang=en\n\nRIKEN AIP will expect adherence to this code throughou t the event. We expect cooperation from all participants to help ensure a safe environment for everybody.\n\n LOCATION:オンライン URL:https://techplay.jp/event/812787?utm_medium=referral&utm_source=ics&utm _campaign=ics END:VEVENT END:VCALENDAR