课件分析machine learning2019s课件auto v

1、MORE ABOUTAUTO-ENCODERHung-yi Lee 李宏毅Auto-encoderAs close as possibleNNEncoderNNDecodervector More than minimizing reconstruction error More interpretable embeddingEmbedding, Latent Representation, Latent CodeWhat is good embedding? An embedding should represent the object.是一對不是一對NNEncoderNNEncoderD

2、iscriminatorimagey/nBeyond Reconstruction binary classifierloss of the classification task is Small The embeddings are representative.Say “Yes”Say “No”How to evaluate an encoder?Train to minimize NNEncoderNNEncoderDiscriminatorimagey/nBeyond Reconstruction binary classifierloss of the classification

3、 task is Small The embeddings are representative.Say “Yes”Say “No”How to evaluate an encoder?Train to minimize Large Not representativeNNEncoderNNEncoderDiscriminatorimagey/nBeyond Reconstruction binary classifierloss of the classification task is Small The embeddings are representative.How to evalu

4、ate an encoder?Train to minimize Large Not representativeTrain to minimize Train the encoder and discriminator to minimize (c.f. training encoder and decoder to minimize reconstruction error)Deep InfoMax (DIM)As close as possibleNNEncoderNNDecodervectorTypical auto-encoder is a special caseNNDecoder

5、vector-score(reconstruction error)DiscriminatorSequential Data Skip thoughtQuick thoughtA document is a sequence of sentences.currentpreviousnextcurrentnextrandomrandomSequential Data Contrastive Predictive Coding (CPC) Auto-encoderAs close as possibleNNEncoderNNDecodervector More than minimizing re

6、construction error More interpretable embeddingEmbedding, Latent Representation, Latent CodeFeature Disentangle An object contains multiple aspect information EncoderDecoderinput audioreconstructed Include phonetic information, speaker information, etc. EncoderDecoderinput sentencereconstructed Incl

7、ude syntactic information, semantic information, etc. Source: Feature Disentangle EncoderDecoderinput audioreconstructed speaker informationphonetic informationEncoder 1Decoderreconstructed speaker informationphonetic informationEncoder 2Feature Disentangle- Voice Conversion How are you?HelloEncoder

8、How are you?EncoderHelloDecoderHow are you?DecoderHelloFeature Disentangle- Voice Conversion How are you?HelloEncoderHow are you?EncoderHelloDecoderHow are you?How are you?Feature Disentangle- Voice Conversion The same sentence has different impact when it is said by different people.Do you want to

9、study a PhD?Go away!StudentStudentDo you want to study a PhD?新垣結衣新垣結衣(Aragaki Yui)Feature Disentangle- Adversarial Training How are you?EncoderHow are you?DecoderHow are you?SpeakerClassifierorLearn to fool the speaker classifier(Discriminator)Speaker classifier and encoder are learned iterativelyFe

10、ature Disentangle- Designed Network ArchitectureHow are you?Encoder 1= instance normalizationINEncoder 2How are you?DecoderINHow are you?(remove global information)How are you?Feature Disentangle- Designed Network ArchitectureHow are you?Encoder 1Encoder 2How are you?DecoderINAdaINHow are you?= inst

11、ance normalizationINAdaIN= adaptive instance normalization(remove global information)(only influence global information)Source Speaker Target Speaker Me (Never seen during training!)MeMeSource to TargetThanks Ju-chieh Chou for providing the results.MeFeature Disentangle - Adversarial Training Discre

12、te Representation Easier to interpret or clustering NNEncoderNNDecoder0.90.10.30.7One-hot1000NNEncoderNNDecoder0.90.10.30.7Binary 1001 11.01144.pdfnon differentiableDiscrete Representation Vector Quantized Variational Auto-encoder (VQVAE) NNEncoderNNDecodervectorvector 1Codebook(a set of vectors)vec

13、tor 2vector 3vector 4vector 5vector 3 For speech, the codebook represents phonetic informationCompute similarity Learn from dataThe most similar one is the input of decoder.Sequence as EmbeddingGRSummary?Seq2seqSeq2seqdocumentdocumentword sequenceOnly need a lot of documents to train the modelThis i

14、s a seq2seq2seq auto-encoder.Using a sequence of words as latent representation.not readable Sequence as EmbeddingGRSeq2seqSeq2seqword sequenceDHuman written summariesReal or notDiscriminatorLet Discriminator considers my output as realdocumentdocumentSummary?ReadableSequence as Embedding Document:澳

15、大利亞今天與澳大利亞今天與13個國家簽署了反興奮劑雙個國家簽署了反興奮劑雙邊協議邊協議,旨在加強體育競賽之外的藥品檢查並共享研究成旨在加強體育競賽之外的藥品檢查並共享研究成果果 Summary: Human:澳大利亞與澳大利亞與13國簽署反興奮劑協議國簽署反興奮劑協議 Unsupervised:澳大利亞加強體育競賽之外的藥品檢查澳大利亞加強體育競賽之外的藥品檢查 Document:中華民國奧林匹克委員會今天接到一九九二中華民國奧林匹克委員會今天接到一九九二年冬季奧運會邀請函年冬季奧運會邀請函,由於主席張豐緒目前正在中南美洲由於主席張豐緒目前正在中南美洲進行友好訪問進行友好訪問,因此尚未決定是否

16、派隊赴賽因此尚未決定是否派隊赴賽 Summary: Human:一九九二年冬季奧運會函邀我參加一九九二年冬季奧運會函邀我參加 Unsupervised:奧委會接獲冬季奧運會邀請函奧委會接獲冬季奧運會邀請函感謝 王耀賢 同學提供實驗結果Sequence as Embedding Document:據此間媒體據此間媒體27日報道日報道,印度尼西亞蘇門答臘島印度尼西亞蘇門答臘島的兩個省近日來連降暴雨的兩個省近日來連降暴雨,洪水泛濫導致塌方洪水泛濫導致塌方,到到26日為日為止至少已有止至少已有60人喪生人喪生,100多人失蹤多人失蹤 Summary: Human:印尼水災造成印尼水災造成60人死亡人死亡 Unsupervised:印尼門洪水泛濫導致塌雨印尼門洪水泛濫導致塌雨 Document:安徽省合肥市最近為領導幹部下基層做了新安徽省合肥市最近為領導幹部下基層做了新規定規定:一律輕車簡從一律輕車簡從,不準搞迎來送往、不準搞層層陪不準搞迎來送往、不準搞層層陪同同 Summary: Human:合肥規定領導幹部下基層活動從簡合肥規定領導幹部下基層活動從簡 Unsupervised:合肥領導幹部下基層做搞迎來送往規定合肥領導幹部