欢迎来到天天文库
浏览记录
ID:40358149
大小:3.15 MB
页数:55页
时间:2019-07-31
《Attention & Memory》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、Attention&Memory集智俱乐部钟翰廷目录●RNNanddynamicsinRNN○FromvanillaRNNtoLSTM/GRU○UnitaryEvolutionRNN○Delay&HierarchiestoReachFather○DynamicsinRNN●AttentionMechanism○Introduction○AttentioninImageCaption/Generation○AttentioninNLP○BetterAttentionMechanism●MemoryNetwork○NeuralTuringMachineandMemory
2、Network○DynamicMemoryNetwork○HierarchicalMemoryNetworkPartI:RNNanddynamicsinRNN●FromvanillaRNNtoLSTM/GRU●UnitaryEvolutionRNN●Delay&HierarchiestoReachFarther●DynamicsinRNNFromvanillaRNNtoLSTM/GRUBasicstructureFromvanillaRNNtoLSTM/GRUBPTT:BackPropagationThroughTimeVanisinggradientproblem
3、FromvanillaRNNtoLSTM/GRUVanisinggradientproblemFromvanillaRNNtoLSTM/GRUPascanu,Razvan,TomasMikolov,andYoshuaBengio."Onthedifficultyoftrainingrecurrentneuralnetworks."ICML(3)28(2013):1310-1318.●Usearegularizationtermthatrepresentsapreferenceforparametervaluessuchthatback-propagatedgradi
4、entsneitherincreaseordecreaseinmagnitude●FromthedynamicalsystemsperspectivewecanseethatpreventingthevanishinggradientproblemimpliesthatwearepushingthemodeltowardstheboundaryofthecurrentbasinofattractionFromvanillaRNNtoLSTM/GRULSTMGRUUnitaryEvolutionRecurrentNeuralNetworks●Whentheeigenv
5、aluesofthehiddentohiddenweightmatrixdeviatefromabsolutevalue1,optimizationbecomesdifficult●uRNNlearnsaunitaryweightmatrix,witheigenvaluesofabsolutevalueexactly1.●uRNNusesReLUasactivationfunction.uRNNCopyingmemoryproblemuRNNAddingProblemDelay&HierarchiestoReachFartherThetemporaldependen
6、ciesarestructuredhierarchicallyDelay&HierarchiestoReachFarther●Delaysandmultipletimescales:Elhihi&BengioNIPS1995,KoutniketalICML2014●HierarchicalRNNs:SordonietalCIKM2015,Mehri&SoroushICLR2017DynamicsinRNNLSTM:ASearchSpaceOdyssey●Theforgetgateandtheoutputactivationfunctionarethecritical
7、componentsoftheLSTMblock●CertainmodificationssuchascouplingtheinputandforgetgatesorremovingpeepholeconnectionssimplifyLSTMwithoutsignificantlyhurtingperformance●LearningrateandnetworksizearethemostcrucialtunableLSTMhyperparameters.Surprisingly,theuseofmomentumwasfoundtobeunimportantD
此文档下载收益归作者所有