BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Chao Wang (UC Davis) DTSTART:20201013T231000Z DTEND:20201014T000000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/1 DESCRIPTION:Title: From telescope to computed tomography via sparse recovery approache s\nby Chao Wang (UC Davis) as part of Mathematics of Data and Decision s at Davis (MADDD) Seminar\n\nAbstract: TBA\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/1/ END:VEVENT BEGIN:VEVENT SUMMARY:Cynthia Rudin (Duke) DTSTART:20201020T231000Z DTEND:20201021T000000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/2 DESCRIPTION:Title: Current Approaches in Interpretable Machine Learning\nby Cynthi a Rudin (Duke) as part of Mathematics of Data and Decisions at Davis (MADD D) Seminar\n\nAbstract: TBA\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/2/ END:VEVENT BEGIN:VEVENT SUMMARY:Patrice Koehl (UC Davis) DTSTART:20201027T231000Z DTEND:20201028T000000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/3 DESCRIPTION:Title: Light speed computation of exact solutions to generic and to degene rate assignment problems\nby Patrice Koehl (UC Davis) as part of Mathe matics of Data and Decisions at Davis (MADDD) Seminar\n\n\nAbstract\nThe l inear assignment problem is a fundamental problem in combinatorial optimiz ation with a wide range of applications\, from operational research to dat a sciences. It consists of assigning ``agents" to ``tasks" on a one-to-one basis\, while minimizing the total cost associated with the assignment. W hile many exact algorithms have been developed to identify such an optimal assignment\, most of these methods are computationally prohibitive for la rge size problems. In this talk\, I will describe a novel approach to solv ing the assignment problem using techniques adapted from statistical physi cs. In particular I will derive a strongly concave effective free energy f unction that captures the constraints of the assignment problem at a finit e temperature. This free energy decreases monotonically as a function of $ \\beta$\, the inverse of temperature\, to the optimal assignment cost\, p roviding a robust framework for temperature annealing. For large enough $\ \beta$ values the exact solution to the generic assignment problem can be derived using a simple round-off to the nearest integer of the elements of the computed assignment matrix. I will also describe a provably convergen t method to handle degenerate assignment problems. Finally\, I will descri be computer implementations of this framework that are optimized for para llel architectures\, one based on CPU\, the other based on GPU. These impl ementations enable solving large assignment problems (of the orders of a f ew 10000s) in computing clock times of the orders of minutes.\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/3/ END:VEVENT BEGIN:VEVENT SUMMARY:Patrick Flaherty (UMass) DTSTART:20201104T001000Z DTEND:20201104T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/4 DESCRIPTION:Title: MAP Clustering under the Gaussian Mixture Model via Mixed Integer P rogramming\nby Patrick Flaherty (UMass) as part of Mathematics of Data and Decisions at Davis (MADDD) Seminar\n\n\nAbstract\nIn the application of clustering models to real data there is often rich prior information th at constrains the relationships among the samples\, or the relationships b etween the samples and the parameters. For example\, in biological or clin ical experiments\, it may be known that two samples are technical replicat es and should be assigned to the same cluster\, or it may be known that th e mean value for control samples is in a certain range. However\, standard model-based clustering methods make it difficult to enforce such hard log ical constraints and may fail to provide a globally optimal clustering. We present a global optimization approach for solving the maximum a-posterio ri (MAP) clustering problem under the Gaussian mixture model. Our approach can accommodate side constraints and preserves the combinatorial structur e of the MAP clustering problem by its formulation as a mixed-integer nonl inear optimization problem (MINLP). We approximate the MINLP through a mix ed-integer quadratic program (MIQP) transformation that improves computati onal aspects while guaranteeing $\\epsilon$-global optimality. An importan t benefit of our approach is the explicit quantification of the degree of suboptimality\, via the optimality gap\, en route to finding the globally optimal MAP clustering. Numerical experiments comparing our method to othe r approaches show that our method finds better optima than standard cluste ring methods. Finally\, we cluster a real breast cancer\ngene expression d ata set incorporating intrinsic subtype information the induced constraint s substantially improve the computational performance and produce more coh erent and biologically meaningful clusters.\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/4/ END:VEVENT BEGIN:VEVENT SUMMARY:Zhi Ding (UC Davis) DTSTART:20201111T001000Z DTEND:20201111T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/5 DESCRIPTION:Title: Deep Learning: Not a Simple Hammer for Massive MIMO Wireless Commun ication Systems\nby Zhi Ding (UC Davis) as part of Mathematics of Data and Decisions at Davis (MADDD) Seminar\n\n\nAbstract\nThe proliferation o f advanced wireless services\, such as virtual reality\, autonomous\ndrivi ng and internet of things has generated increasingly intense pressure to\n develop intelligent wireless communication systems to meet networking need s\nposed by extremely high data rates\, massive number of connected device s\, and ultra\nlow latency. Deep learning (DL) has been recently emerged a s an exciting design\ntool to advance the development of wireless communic ation system with some\ndemonstrated successes. In this talk\, we introduc e the principles of applying DL for\nimproving wireless network performanc e by integrating the underlying\ncharacteristics of channels in practical massive MIMO deployment. We develop\nimportant insights derived from the p hysical RF channel properties and present a\ncomprehensive overview on the application of DL for accurately estimating channel\nstate information (C SI) of forward channels with low feedback overhead. We\nprovide examples o f successful DL application in CSI estimation for massive MIMO\nwireless s ystems and highlight several promising directions for future research.\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/5/ END:VEVENT BEGIN:VEVENT SUMMARY:Chelsea Weaver (Amazon) DTSTART:20201118T001000Z DTEND:20201118T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/6 DESCRIPTION:Title: Natural Language Understanding at Amazon Music\nby Chelsea Weav er (Amazon) as part of Mathematics of Data and Decisions at Davis (MADDD) Seminar\n\n\nAbstract\nIn this talk\, I’ll discuss what happens when you ask an Alexa device to play music. I’ll focus on the Natural Language U nderstanding (NLU) component\, which deals with categorizing and labeling transcribed requests. In particular\, I’ll discuss two projects I’ve w orked on designed to improve upon the initial labeling. The first uses a B ERT-based language model to “correct” requests that appear to be misla beled. The second is an online learning model that selects from different NLU interpretations using implicit customer feedback. I’ll conclude the talk with a few tips for the industry job search.\n\nLinks: Amazon Jobs Pa ge\; Amazon Science Page\n\nPapers:\nPersonalizing natural-language unders tanding using multi-armed bandits and implicit feedback – Moerchen et al (2020)\nCounterfactual Risk Minimization: Learning from Logged Bandit Fee dback - Swaminathan & Joachims (2015)\nAnalysis of Thompson Sampling for the Multi-Armed Bandit Problem – Agrawal et al (2012)\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/6/ END:VEVENT BEGIN:VEVENT SUMMARY:Wolfgang Polonik (UC Davis) DTSTART:20201125T001000Z DTEND:20201125T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/7 DESCRIPTION:Title: Multiscale Geometric Feature Extraction\nby Wolfgang Polonik (U C Davis) as part of Mathematics of Data and Decisions at Davis (MADDD) Sem inar\n\n\nAbstract\nA method for extracting multiscale geometric features from a data cloud is presented. Each pair of data points is mapped into a real-valued feature function\, whose construction is based on geometric co nsiderations. The collection of these feature functions is then being used for further data analysis. Applications include classification\, anomaly detection and data visualization. In contrast to the popular kernel trick\ , the construction of the feature functions is based on geometric consider ations. The performance of the methodology is illustrated through applicat ions to real data sets\, and some theoretical guarantees supporting the pe rformance of the novel methodology are presented. This is joint work with G. Chandler.\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/7/ END:VEVENT BEGIN:VEVENT SUMMARY:Guido Montufar (UCLA) DTSTART:20201202T001000Z DTEND:20201202T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/8 DESCRIPTION:Title: Optimal Transport to Independence Models\nby Guido Montufar (UC LA) as part of Mathematics of Data and Decisions at Davis (MADDD) Seminar\ n\n\nAbstract\nAn independence model for discrete random variables is a Se gre-Veronese variety in a probability simplex. Any metric on the set of jo int states of the random variables induces a Wasserstein metric on the pro bability simplex. The unit ball of this polyhedral norm is dual to the Lip schitz polytope. Given any data distribution\, we seek to minimize its Was serstein distance to a fixed independence model. The solution to this opti mization problem is a piecewise algebraic function of the data. We compute this function explicitly in small instances\, we examine its combinatoria l structure and algebraic degrees in the general case\, and we present som e experimental case studies. This talk is based on joint work with Türkü Özlüm Çelik\, Asgar Jamneshan\, Bernd Sturmfels\, Lorenzo Venturello.\ n\nhttps://arxiv.org/abs/1909.11716\n\nhttps://arxiv.org/abs/2003.06725\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/8/ END:VEVENT BEGIN:VEVENT SUMMARY:Samir Chowdhury (Stanford) DTSTART:20201209T001000Z DTEND:20201209T010000Z DTSTAMP:20260422T212753Z UID:MADDD_Fall2020/9 DESCRIPTION:Title: Gromov-Wasserstein Learning in a Riemannian Framework\nby Samir Chowdhury (Stanford) as part of Mathematics of Data and Decisions at Davi s (MADDD) Seminar\n\n\nAbstract\nGeometric and topological data analysis m ethods are increasingly being used to derive insights from data arising in the empirical sciences. We start with a particular case where such techni ques are applied to human neuroimaging data to obtain graphs which can the n yield insights connecting neurobiology to human task performance. Reprod ucing such insights across populations requires statistical learning techn iques such as averaging and PCA across graphs without known node correspon dences. We formulate this problem using the Gromov-Wasserstein (GW) distan ce and present a recently-developed Riemannian framework for GW-averaging and tangent PCA. Beyond graph adjacency matrices\, this framework permits consuming derived network representations such as distance or kernel matri ces\, and each choice leads to additional structure on the GW problem that can be exploited for theoretical and/or computational advantages. In part icular\, we show how replacing the adjacency matrix representation with a spectral representation leads to theoretical guarantees allowing efficient use of the Riemannian framework. Additionally we present numerics showing how the spectral representation achieves state of the art accuracy and ru ntime in graph learning tasks such as matching and partitioning on a varie ty of real and simulated datasets.\n LOCATION:https://researchseminars.org/talk/MADDD_Fall2020/9/ END:VEVENT END:VCALENDAR