Abstract：WewilltalkaboutarecentworkonthestabilityofnilpotentstructuresonacollapsedmanifoldwithboundedsectionalorRiccicurvature.Amanifoldofboundedsectionalcurvatureiscalledepsilon-collapsed,iftheinjectivityradius,orequivalentlythevolumeofunitball,ateverypointsislessthanepsilon.Thegeometry/topologyofacollapsedmanifoldcanbetotallydescribedbyCheeger-Fukaya-Gromov\'snilpotentstructure.SimilarresultshadbeenextendedtomanifoldsofboundedRiccicurvatureundersomeadditionalassumptions.ThestabilityoflocallydefinednilpotentstructureswereessentialintheworkofCheeger-Fukaya-Gromovtoconstructaglobalnilpotentstructureononefixedmetric.Nilpotentstructuresalsodependonthechoiceofepsilon,thecollapsedlengthscaleoneinspects.WeprovethatiftwometricsonamanifoldareL_0-bi-Lipchitzequivalentandsufficientcollapsed(dependingonL_0)underbounded(Ricci)curvature,thentheunderlyingnilpotentstructuresareisomorphictoeachother.Asapplications,weestablishalinkbetweenthecomponentsofthemodulispaceofallcollapsedRiemannianmetricsandthesetofisomorphismclassesofnilpotentstructures,andderiveanewparametrizedversionofGromov\'sflatmanifoldtheoremunderboundedRiccicurvatureandconjugateradius.

报告人简介:陈凌，女，中国科学院地质与地球物理研究所研究员，博士生导师，国家杰出青年基金获得者（2012年）、科技部中青年科技创新领军人才（2015年度）。1994年毕业于中国科学技术大学地球与空间科学系，2002年获中国地震局地球物理研究所固体地球物理专业博士学位。1998、2000-2002年美国加州大学SantaCruz分校访问学者，2010年5-11月麻省理工学院访问教授。主要从事地震偏移成像方法、流动地震台阵观测和壳幔结构高精度地震成像研究，并通过多学科观测结合，来认识岩石圈和上地幔性质、状态和动力学，以及大陆演化。2005年以来主持了华北、华南和伊朗地区多项密集流动地震台阵探测工作。已发表SCI论文50余篇，SCI他引上千次。获中国科学院杰出科技成就奖、王宽诚教育基金会优秀女科学家专项奖、中国地球物理学会傅承义青年科技奖等荣誉。现任《中国科学地球科学》副主编、《地质科学》副主编，《SolidEarthSciences》编委。

Bio：SongHangraduatedfromStanfordUniversityadvisedbyProf.BillDally.HewilljoinMITEECSasassistantprofessorinJuly2018.Hisresearchfocusesonenergyefficientdeeplearning,attheintersectionbetweenmachinelearningandcomputerarchitecture.HeproposedtheDeepCompressionalgorithm,whichcancompressneuralnetworksby189×whilefullypreservingpredictionaccuracy.Hedesignedthefirsthardwareacceleratorthatcanperforminferencedirectlyonacompressedsparsemodel,whichresultsinsignificantspeedupandenergysaving.HisworkhasbeenfeaturedbyO’Reilly,TechEmergence,TheNextPlatform,andEmbeddedVision,andithasimpactedtheindustry.HeledresearcheffortsinmodelcompressionandhardwareaccelerationthatwontheBestPaperAwardatICLR’16andtheBestPaperAwardatFPGA’17.BeforejoiningStanford,SonggraduatedfromTsinghuaUniversity.Abstract：DeeplearninghasspawnedawiderangeofAIapplicationsthatarechangingourlives.However,deepneuralnetworksarebothcomputationallyandmemoryintensive.Thustheyarepowerhungrywhendeployedonembeddedsystemsanddatacenterswithalimitedpowerbudget.Toaddressthisproblem,Iwillpresentanalgorithmandhardwareco-designmethodologyforimprovingtheefficiencyofdeeplearning.Iwillfirstintroduce"DeepCompression",whichcancompressdeepneuralnetworkmodelsby18-9×withoutlossofpredictionaccuracyforabroadrangeofCNN,RNN,andLSTMs.Thecompressionreducesbothcomputationandstorage.Next,bychangingthehardwarearchitectureandefficientlyimplementingDeepCompression,IwillintroduceEIE,theEfficientInferenceEngine,whichcanperformdecompressionandinferencesimultaneously,savingasignificantamountofmemorybandwidth.Bytakingadvantageofthecompressedmodelandbeingabletodealwithanirregularcomputationpatternefficiently,EIEachieves13×speedupand3000×betterenergyefficiencyoverGPU.Finally,Iwillrevisittheinefficienciesincurrentlearningalgorithms,presentDSDtraining,anddiscussthechallengesandfutureworkinefficientdeeplearning.Attheendoftheseminar,IwillintroducethePh.D.recruitmentintheareaofartificialintelligenceandcomputerarchitectureatMIT.

意大利卓越计划宣讲会暨粒子物理讲座时间：2017年9月18日（周一）上午10:30-12:00地点：近代物理学院210会议室主讲人：1、PROF.MARCOMAGGIORA，DirectoroftheIHEP-INFNJointLaboratory；ScientificAssociateTurinSection,NationalInstituteofNuclearPhysics(INFN)，AssociateProfessorDepartmentofPhysics,UniversityofTurin2、PROF.FABRIZIOBIANCHI,ResearcherofScientificAssociateTurinSection,NationalInstituteofNuclearPhysics(INFN)3、PROF.CARLOD’IMPORZANO:ItalianExcellenceProgramGeneralDirector,professorandRectorofUniversity,formerCNRassociateresearcheronalgebraicstructuresofthelanguage内容：1、中国和意大利粒子物理领域的科研合作；2、IHEP-INFNJointLaboratory项目，以及人才培养需求；3、都灵大学与中科大的项目合作以及项目合作前景；4、意大利卓越计划物理领域的学术网络，以及计划的国际化项目（优秀人才培养规划）等。

摘要：Inthistalk,weshallestablishtheasymptoticlog-Harnackinequalityforarangeofpath-dependentSDEs,whichallowthelengthofmemorytobeinfiniteandthediffusiontermstobedependentfullyonthepasthistory.Werevealthattheasymptoticlog-Harnackinequalityimplies(i)theasymptoticheatkernelestimate,(ii)theabsolutecontinuityofthetransitionkernelw.r.t.theinvariantprobabilitymeasureand(iii)theuniquenessofinvariantprobabilitymeasure(ifitexists).Asanotherbyproductoftheasymptoticlog-Harnackinequality,wealsoderivetheasymptoticgradientestimate,whichfurtherimpliesthatthesemigroupgeneratedbythesegmentprocessenjoystheasymptoticstrongFellerproperty.

2017年秋季课程

Abstract:In1947MarkKacprovedthatforacentredHoeldercontinuousfunction$f$definedon$[0,1]$andextendedperiodicallytothewholerealline,thesequence$f(2^kt)$satisfiesaCentralLimitTheorem.Weextendtheresulttomeasurable$L^2$functionssatisfyingafastapproximationproperty.TheresultnaturallyextendstosimilarfunctionsdefinedonBernoullishifts.

地空学院杰出讲座的报告通知报告地点:东区水上报告厅报告时间:2017-09-1815:00报告人:Dr.L.RubyLeung报告人简介:美国工程院院士，美国AAAS,AGU,和AMS会士报告题目:全球变暖导致极端天气事件频发？

摘要：Thistalkfocusesonexploringtheglobalstabilityofnonlinearstochasticfeedbacksystemsonthenonnegativeorthantdrivenbymultiplicativewhitenoiseandpresentingacoupleofsmall-gainresults.Weinvestigatethedynamicalbehaviorofpull-backtrajectoriesforstochasticcontrolsystemsandprovethatthereexistsauniquegloballyattractingpositiverandomequilibriumforthosesystemswhoseoutputfunctionseitherpossessboundedderivativesorareuniformlyboundedawayfromzero.Ourresultscanbeappliedtowell-knownstochasticGoodwinnegativefeedbacksystem,Othmer-TysonpositivefeedbacksystemandGriffithpositivefeedbacksystemaswellasotherstochasticcooperative,competitiveandpredator-preysystems.ThisisajointworkwithDr.LvXiang.