计算图与反向传播：理解梯度计算.pptxVIP

ComputingGradientHung-yiLee李宏毅

IntroductionBackpropagation:anefficientwaytocomputethegradientPrerequisiteBackpropagationforfeedforwardnet:Simpleversion:BackpropagationthroughtimeforRNN:(v6).ecm.mp4/index.htmlUnderstandingbackpropagationbycomputationalgraphTensorflow,Theano,CNTK,etc.

ComputationalGraph

ComputationalGraphA“language”describingafunctionNode:variable(scalar,vector,tensor……)Edge:operation(simplefunction)?abcExample????xuvy????

ComputationalGraphExample:e=(a+b)?(b+1)abcdee=c?dc=a+bd=b+1?++121326

Review:ChainRuleCase1