Why Momentum Really Works

Why Momentum Really Works

ID:41501271

大小:2.31 MB

页数:19页

时间:2019-08-26

Why Momentum Really Works_第1页
Why Momentum Really Works_第2页
Why Momentum Really Works_第3页
Why Momentum Really Works_第4页
Why Momentum Really Works_第5页
资源描述:

《Why Momentum Really Works》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、WhyMomentumReallyWorksStartingPointOptimumSolutionStepsizeα=0.02Momentumβ=0.99WeoftenthinkofMomentumasameansofdampeningoscillationsandspeedinguptheiterations,leadingtofasterconvergence.Butithasotherinterestingbehavior.Itallowsalargerrangeofstep‑sizestobeused,and0

2、0.0030.0060.000.5000.990createsitsownoscillations.Whatisgoingon?GABRIELGOHUCDavisApril.42017Citation:Goh,2017Here’sapopularstoryaboutmomentum[1,2,3]:gradientdescentisamanwalkingdownahill.Hefollowsthesteepestpathdownwards,hisprogressisslow,butsteady.Momentumisahea

3、vyballrollingdownthesamehill.Theaddedinertiaactsbothasasmootherandanaccelerator,dampeningoscillationsandcausingustobarrelthroughnarrowvalleys,smallhumpsandlocalminima.Thisstandardstoryisn’twrong,butitfailstoexplainmanyimportantbehaviorsofmomentum.Infact,momentumc

4、anbeunderstoodfarmorepreciselyifwestudyitontherightmodel.Onenicemodelistheconvexquadratic.Thismodelisrichenoughtoreproducemomentum’slocaldynamicsinrealproblems,andyetsimpleenoughtobeunderstoodinclosedform.Thisbalancegivesuspowerfultractionforunderstandingthisalgo

5、rithm.WebeginwithGradientDescent.Thealgorithmhasasmanyvirtues,butspeedisnotoneofthem.Itissimple--whenoptimizingasmoothfunctionf,wemakeasmallstepinthegradientwk+1=wk−α∇f(wk).Forastepsizesmallenough,gradientdescentmakesamonotonicimprovementateveryiteration.Italways

6、converges,albeittoalocalminimum.Andunderafewweakcurvatureconditionsitcanevengetthereatanexponentialrate.Buttheexponentialdecrease,thoughappealingintheory,canoftenbeinfuriatinglysmall.Thingsoftenbeginquitewell--withanimpressive,almostimmediatedecreaseintheloss.But

7、astheiterationsprogress,thingsstarttoslowdown.Youstarttogetanaggingfeelingyou'renotmakingasmuchprogressasyoushouldbe.Whathasgonewrong?Theproblemcouldbetheoptimizer'soldnemesis,pathologicalcurvature.Pathologicalcurvatureis,simplyput,regionsoffwhicharen'tscaledprop

8、erly.Thelandscapesareoftendescribedasavalleys,trenches,canalsandravines.Theiterateseitherjumpbetweenvalleys,orapproachtheoptimuminsmall,timidsteps.Progressalon

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。