19
play

19 Auto Lecture encoders : Ankur Bambhanoliya Scribes : - PowerPoint PPT Presentation

Feb 22, 2024 •444 likes •673 views

I Variational 19 Auto Lecture encoders : Ankur Bambhanoliya Scribes : Donald Hamnett Motivation Inferring from Images Latent Variables : Dataset MNIST Goh images ; hand of digits written - Goal two variables Infer

  1. I Variational 19 Auto Lecture encoders : Ankur Bambhanoliya Scribes : Donald Hamnett

  2. Motivation Inferring from Images Latent Variables : Dataset MNIST Goh images ; hand of digits written - Goal two variables Infer labels 1. Digit , Style { 9 } y e 0 - , . . , variables 2 RD 7 €

  3. Assume all digits Deep IX g Generative Models frequent equally , ) Neural Network Idea µyn,7n I Use to define : a , mode generative a generative µ .ae ( Digit ) supervise can ( to ) Discrete yn ~ 0.1 0.1 . . . , , Normal I ) ( 2- a n n , Berinoulli ( :O ) ) Xn n I Neural network Questions of 9 How ? Style train I this do we . ? supervision ) How inference do ( Image do 2 no we .

  4. Generative Training Models Deep Gradient Use Ascent Idea Stochastic 2 : on a lower bound approximate likelihood estimation to maximum Generative model with True data weights O network distribution a > KL ( pdatacxsllpcxio ) - To ) I T leg E = - paan , pcxnlo It . xhnpdatac :3 ftp.logpcxioyfFEIFstiaip Epdata = from unknown a distribution data N I , § To log , ) = = ,

  5. Generative Training Models Deep Gradient Use Ii II Ascent Idea Stochastic 2 : 10 on a lower bound approximate likelihood estimation to maximum .HN//ply,z1x,oD)/ variational distribution I proposal depends x on : ' " Ll o ' ' ftp.nm.lt#gI..leos' KL ( 9cg panta ( hog E ) PK = - Tx - Cx ) 's , I Ipo - To ( pm KL JL Col qcyitlxl ply 7190 when ) - = - , ] ) # I Ealy , ,×,[ To log 0£10 ) Ftpdnta 7107 pix , b. = pdatacx , n Tuff ? x - leg Do pcxn.gs , # = to , 5,7 ' 71×4 racy ,

  6. Generative Training Models Deep Gradient Use ;D ÷ : Ascent Idea Stochastic 3 : Learn Variational perform inference to doin Ii : " 't . pan .am#as.....teost Perform Combining 1+2 gradient both i ascent on Max To 7 , 4) Likelihood I ( O : . £10,4 model ) OF ( leg generative pgcxl wgnfax Variational Inference Tq ) 2 : . ) ) ( Approx . Posterior ) 't arqninklfqcyit.gs/lpry,7Ix of =

  7. Generative Training Models Deep Use Idea 4 neural network to define : ( . ) inference variation model the a.k.a. the dist Digit Inference model f I '(xn ;¢↳ ) ) ( Discrete yn ~ networks } Neural Normal ( Itkmyn ;¢D 7 n ~ y.tl/)=Mnqcyn,7n1xn18tule Vaniaticnal Dtst / in q( Image

  8. Variational Auto encoders Learn Objective model generative deep and : a a by inference corresponding model optimizing Em ) Lead Oo . F. ' - - . .mn/logP:Yi;::iIiTI -

  9. Auto Intermezzo encoders : 7h Xn ( all continuous >

  10. latent Encoder Mapping from code to image : x 7 Multi Perception 768 256 whijxn bh 50 layer 2- - hn ;=6( { ;) , ; + . wtinhni - Activation linear map o ( § 10h Tainan Qtnptoochws bti ) + zn ;= , , Activation functions 6kt X

  11. Mapping Decoder from to code image 7 x : Multi Perception layer . • ( § ) BY wiijzn hni + = , ; Whinhnj gxnzooh dipterous + bhi ) pwws = 6( § In . ; Loss Binary - entropy Cross : ,¢)=n§ .pl?jxn.plogxn.p LH ^( Minimize SGD with =

  12. Learned Auto Latent Codes encoder

  13. Auto Variation Treat latent al variable encoder a : as

  14. Auto Variational Treat latent variable encoder a : as Encoder ) Model ( ( Decoder ) Inference Model Generative O ) ph , 7- Q ) 917=1 Xn pin ; 's n - g 6 ( Whxn tbh ) ( Normal I ) hn 7- o = n n , ¢M Oh 6 ( Whan → on → , tbh ) " " µ ? b hut hn W = = 46 ' I ok - r a of What " Whn Git ) exp I y :-. b b = + ( pint , 67 ) Normal ( Yn ) Bernoulli 7- n xn n n 7×71 , on I log Pg ' , ] Lto Objective Eg , on : - - , ×

  15. Variutional Auto Learned Latent Codes encoder

  16. Variation Auto Auto encoder encoder vs

  17. Training The Re Trick parameterization : fowl qczb lxb tf batch Compute of gradient images : on % " Pok " Uniform ( f X xbn . ,Xr3 ) ) gczhsix , , . . B ¢ .iq#gczixbsllogqoTzx Analogue BBVI REINFORCE Estimator of style : - - § , 's § To b b s , to , LIO ) log , a) ' ' , log = 's ) B a Y b. s b 2- ( 2- 7 I X ~ g ¢ Problem will be high variance very ;

  18. Training The Re Trick parameterization : = Mkt = toff eh Eb b Idea Sample 's parameterized dist z using a : te +61×3 b. s Normal } ( 0,1 ) q ~ b. s ebiipce b it 9ft ) Ix ~ s :p ) ;¢ , 7 Tret ^ Neural works Result B- Estimator Re parameterized pce.bg/b8qo,czces;qslxb I Poet : b ;p ) ) fo I z £10,10 ) , Tq ) . B b b. s s ) ' 0 ) " E Po " [ Is ' leg , = 4 b. s B ; 9) 94 ( 2- C E lxb ) be s - I - . In S often I enough practice :

  19. Variational Auto encoders Learn Objective model generative deep and : a a by inference corresponding model optimizing ) scores % F. - - .GE?Em..*.leogP:YiiiiiiiTI -

  20. ( 7 style & digit ) Continuous ,7n ) ( Xn both pp nlxnl 9$17 : encodes , ft style ) Continuous Discrete ynifnl pfxn encodes qfbn lxn ) : ,7n + , ,

  21. Disentangled Representations Hao Learn Interpretable Babahtsarthaht Features : Xn Zn ' , Zn , -2 : in ,7 C 3 3 - ÷ 1 1 ) × 6 Fn ,i Zn 7h ,i=3 =-3 ,i=o

Recommend

More recommend