于恒RNN和LSTM神经网络的原理及应用.pdf
于恒-RNN和LSTM神经网络的原理及应用云栖大讲堂-编程语言专场 Recent breakthrough in Machine Translation知识就是力However, the sky remained clear under the strong north windEncoder4However the sky remained clear under the strong north wind[(2)虽約匚但医空依然十分清敝[北呼啸Decoddddd虽然[北风呼啸。但天空依然十分清澈□。Although north wind howls but sky still extremely limpidKnowledgeispowercendsDefine smarterTomorrowOutline∠2RNN原理■RNNs的“不讲理有效性Network architecturea Image captionAttention mechanismQuestion AnsweringExampleMachine translation既生RNN,何生LSTM?RNNs的阿喀琉斯之踵在哪里?Advanced architecturerainingOther rnn variantsDefine smarterTomorrowRNN原理 Network structure∠2Motivation: Process arbitrary length of inputTTEr(七CONTEXT (C)W○-.○-○t+IUnfoldUUxArecurrent neural network and the unfolding in time of the computation involved in itsforward computation. Source: NatureCONTEXT (C-1)Define smarterTomorrowRNN原理 Network structure∠2We can process a sequence of vectors x byapplying a recurrence formula at every time stepht=Ifw(ht-l,atRNNnew stateold state input vector atsome time stepsome functionwith parameters wDefine smarterTomorrow6RNN原理 Network structure∠2yht= fw(ht-1,t)RNNht= tanh(Wrhht-1+WhatytWha,hyDefine smarterTomorrowRNN原理 Computation Graph∠2hXDefine smarterTomorrowRNN原理 Computation Graph∠202Define smarterTomorrowRNN原理 Computation Graph∠2每个共用同一组权重矩阵23W2310
下载地址
用户评论
