11-785

CMU 11-785 L12 Back propagation through a CNN

ConvolutionEachpositioninzzzconsistsofconvolutionresultinpreviousmapWayforshrinkingthemapsStridegreaterthan1Downsampling(notnecessary)Typicallyperformedwithstrides>1PoolingMaxpoolingNote:keeptrackingo
zealscott·2020-08-19 01:06

CMU 11-785 L16 Connectionist Temporal Classification

SequencetosequenceSequencegoesin,sequencecomesoutNonotionof“timesynchrony”betweeninputandoutputMayevennotsmaintainorderofsymbols(fromonelanguagetoanother)WithordersynchronyTheinputandoutputsequencesha
zealscott·2020-08-19 00:48

CMU 11-785 L17 Seq2seq and attention model

GeneratingLanguageSynthesisInput:symbolsasone-hotvectorsDimensionalityofthevectoristhesizeofthe「vocabulary」Projecteddowntolower-dimensional“embeddings”Thehiddenunitsare(oneormorelayersof)LSTMunitsOutp
zealscott·2020-08-19 00:48

CMU 11-785 L13 Recurrent Networks

ModellingSeriesInmanysituationsonemustconsideraseriesofinputstoproduceanoutputOutputstoomaybeaseriesFiniteresponsemodelCanuseconvolutionalneuralnetappliedtoseriesdata(slide)AlsocalledaTime-Delayneural
zealscott·2020-08-19 00:47

CMU 11-785 L15 Divergence of RNN

VariantsonrecurrentnetsArchitecturesHowtotrainrecurrentnetworksofdifferentarchitecturesSynchronyThetargetoutputistime-synchronouswiththeinputThetargetoutputisorder-synchronous,butnottimesynchronousOne
zealscott·2020-08-19 00:47

CMU 11-785 L14 Stability analysis and LSTMs

StabilityWillthisnecessarilybe「BoundedInputBoundedOutput」?GuaranteedifoutputandhiddenactivationsareboundedButwillitsaturate？AnalyzingRecursionSufficienttoanalyzethebehaviorofthehiddenlayersinceitcarri
zealscott·2020-08-19 00:47
上一页 1 下一页
按字母分类： ABCDEFGHIJKLMNOPQRSTUVWXYZ其他