Lock in $30 Savings on PRO—Offer Ends Soon! ⏳

DeNA, MoT合同輪講発表資料 / domain adaptation

DeNA, MoT合同輪講発表資料 / domain adaptation

2020.07.02 DeNA, Mobility Technologies合同の勉強会にて発表に使用した資料です。

Domain Adaptationってなにか、どんな考え方、手法があるのかについて紹介しました。

Takumi Karasawa

July 02, 2020
Tweet

More Decks by Takumi Karasawa

Other Decks in Research

Transcript

  1. ) , . , 3 gT s D d p

    D n ( h t M D ic ( ,, o C D b VL M V l e D 
  2. - . , , , A , , P ,

    , N pyu gT NdCbaceI Apyu TskD A pyu whD , L A , , F if AM m pyu o lt n L C r 6 DATASET 1 DATASET 2 knowledge SYSTEM 1 SYSTEM 2
  3. . , CL M D F ! = #, %

    & n # n % & , 'ℎ)*) & = +, , … , +. ∈ # 0 = {2, % 3|& } n 2 n % 3|& , 'ℎ)*) 3 = 6, , … , 6. ∈ 2 , , , 7
  4. , . , D L C A S TA S

    T M L , S T S T , S T 8          
  5. , /- / . . Fca S chi m n

    t Duhi m b a -/Ahi m hi m c eg / Ahi m o c eghi m a ed y / F l FA tC FTb ahi m y l c eghi m p . AM F ➤ ca y s b a c c eg c c eghi m l L C r 9
  6. - . - , T - -, cde hUb cU

    T t C de hTmM Tl b iL r T bgu M p v U n - -, Uoy“ cb n - - -, Uoy“ cb oySM cb n - -, A UoySM cb D C sn M C U T T U / - / - / - - 10
  7. -3 3 .2 13 , , ,.2 - o m

    s , ,.2 l r - - - - - e t ,- 3 1 , ,.2 n - - ai n - - I - e t  12 70 20 10 g
  8. . , .,. . M C C T L -

    13   " !    # !$ "
  9. 1 A 42 4 3 t gdRNC RT M z

    t b M L mMC S L O . 2 4 C ,- D d r my 3 4 1 43Dusvcn -42 2 1 43Dlioh 1 43D ea - 14
  10. ., . - he AMT R n cg nOdM iObl

    CD o n a M L M n .- n 15
  11. , , . , t gi yd lg !" c

    T L M s i n !# C h n lg n n t i t o n L b b L D n b L e 16    
  12. , , . , n C N n M L

    n D A n 17 Classification loss domain loss M A DDC [1]
  13. . . - ,. n . n 5K 5 M

    ➤ 5 M 8 n DT D K L n T C K TKM n 3 4 n 3 4 1 - n 4 6 18 DAN 5C DF [2]
  14. . , . Dg e l D Cc n a

    C M D dl L T dlC M D dl h L C h D n ib 20 (a) ➤ (b) ➤ (c)CORAL  (d)    [4] ➤ (b) ➤ (c) CORAL  [4]
  15. , M T LRD O n LR C C n

    . C A , 21  Covariance matrix % ("$ fc8 *# '+ & ) [5] Classification loss domain loss !  i  
  16. . , .,. . L L T M C L

    L C M 22          
  17. . . ,. L G e T . i a

    b L G M h R n . d Cd g c 23 DANN [6]
  18. . 2 . nbTd S T L [ Tg d

    toc h E [ C sr le . , 1 Td Mf [ [ C , . , T A ] aMsr leTg S n A A n A n sr le d i . , 1 T A m 24 ADDA [7]
  19. . 25 M geh a hM M J i Tl

    oS b GJ L C LDK , . A dn N S GJec G N !"# = %&'()*(+)||. + )/%&'()0(+)||. + ) 1 2 3 = )*(+)/)0(+) 1 ➤ 2!"# = 5)6 log )*(+) )* + /)0(+) + 5)0 log )0(+) )* + /)0(+) + log 4 5+6~=* log ! ># 3# + 5+0~=0 log 1 − ! >A 3A
  20. . , . C LM D Dab TC . C

    - .1 - - . - .3 26    
  21. . - leCgc f N T L h a i

    gc b , N T d n L mCgc b o M p n N T , C n 27 DRCN  [8] Reconstruction loss Classification loss
  22. . , .,. . TM n ic TM t icL

    L s TM t Cn L n l TM t C TM n h gb !"#$$%&%'(% n TM TM n d gb !)#*#+,&#-. n L o L e !&%(/' 28 DSN  [9]
  23. . , .,. . C T M C L L

    C - . 29 [10] 
  24. ,2 2 1 2 . b 1 2 2 T

    e d c a][ NL M C B 0 1 30
  25. ,2 2 1 2 . C Mv tas n BLe

    N b M Dn Mc o T d T T Mr C 2 Mbg m i N b Mh A l Mbg 0 1 31 AdaBN  [10]
  26. . ed , c a N L C T C

    M I Cb 1 1 1 32 BN  IN   [Y. Wu+, 2018] Style transfer  [11] Content Style
  27. , . - , ovc -, -, gr I P

    T LM h C T L t dm n , , , e i asn , , L bl n M T - . 33           
  28. 2 2 . , N u Nl MnG Nn G

    g r , ci o N b T N j sd eA LC D M tu L h 2 1 2 34 CoGAN  [12]  [12]
  29. . , ba L D c e n L d

    A N G T hgC n M i DC . 35 DDiscriminator GGenerator TTask-specific model task-specific loss domain loss (GAN loss) PixelDA  [13]
  30. + 17 7 - 2 7 ui0 2 , .

    0 n AG n [ ! " AG e #$ #% MN , bo n 2 2 7 2 s n [ AZc dLt n yN ] CN T ] h [ g n ] l 36 Cycle-consistency loss  [Zhu+, 2017]
  31. - . . , , , , , A .

    D L NC T M D , , ,. G n FD , . . . S 37 !"#"$%$ = !'()* + !,$-. + !,$-/ + !0#0 + !,$-1 + !)23 Cycle GAN     [14] 
  32. . , .,. . LC T M - .. -

    40    {"# , % &' } "# % &'  
  33. - - , - . - -, , n enL

    ml c ! !" , !$ sL !% n !" , !$ MU T dh rt gu bi n !" , !$ L Cpa voCfA DL dh rt n !% Mdh rt L uA bi 41 ATDA  [15]
  34. . , .,. .  42 MMD DDC(2014) DAN(2015) JAN(2017)

    Adversarial DANN(2015) ADDA(2017) Reconstruction DRCN(2016) DSN(2016) Normalization AdaBN(2017) AdaIN(2017) Input-level CoGAN(2016) PixelDA(2017) CyCADA(2018) Pseudo-labelling ATDA(2017) CORAL CORAL(2016) DeepCORAL(2016)
  35. 5 CJMJTY :FDIO M HJFS , 4TE A FF NBJO

    , O SJ O[ , . :ZFOH 2 0 NBO 6 IBOH 3 BFOL BOE : BRRFMM FF E NBJO D O SJ O 5BXJNJZJOH R E NBJO JOVBRJBODF 1O BR JV A FF EB TBTJ O 6FTW RL[ 6 5 4 OH ,B 2 BOH BOE 5 2 REBO 4FBROJOH TRBOS FRBCMF FBT RFS WJTI EFF BEB TBTJ O OFTW RLS 1O 1,54 A 2 JOT EB TBTJ O 6FTW RL[2 6 5 4 OH 2 BOH BOE 5 1 2 REBO FF TRBOS FR MFBROJOH WJTI K JOT BEB TBTJ O OFTW RLS 1O 1,54 A ,7 RFMBTJ O 4JHONFOT[,7 4 O 2 FOH BOE 3 BFOL FT RO R STRBTJOHMY FBSY E NBJO BEB TBTJ O 1O 1 A FF ,7 4 O BOE 3 BFOL FF D RBM , RRFMBTJ O BMJHONFOT R EFF E NBJO BEB TBTJ O 1O .,, W RLSI 43
  36. 5 DKNKUY GEJP N IKGT , U A B -

    CKP WGSTCSKCN 6GVSCN 6GUX SMT[- 66 CPKP CP G RKUTMY PTVRGSWKTG CKP C CRUCUK P DY DCEMRS RCICUK P 1P 1,5 A B WGSTCSKCN -KTESK KPCUKWG - CKP CRUCUK P[ -- . ZGPI 2 0 HH CP 9CGPM CP -CSSGNN WGSTCSKCN KTESK KPCUKWG CKP C CRUCUK P 1P , 78 A B -GGR 8GE PTUSVEUK P ,NCTTKHKECUK P 6GUX SM[-8,6 5 JKHCSY NGKLP 5 JCPI - CN VZZK CP K -GGR SGE PTUSVEUK P ENCTTKHKECUK P PGUX SMT H S VPTVRGSWKTG CKP C CRUCUK P 1P .,, A B - CKP 9GRCSCUK P 6GUX SM[-96 VT CNKT SKIG SIKT 6 9KNDGS CP - SKTJPCP CP - .SJCP - CKP TGRCSCUK P PGUX SMT 1P 6GVS179 A B CRUKWG CUEJ 6 S CNKZCUK P[ C 6 K 6 CPI 2 9JK 2 KV CP 0 V 8GWKTKUKPI DCUEJ P S CNKZCUK P H S RSCEUKECN CKP C CRUCUK P 1P 1, 8 X SMTJ R 44
  37. 7R MOMV[ IGL RORKMIU .R 6VH C D ,HESVMXI 3

    UVE GI 8RTPEOM EVMR ],HE38 2 E K E H -IOR KMI ,T MVTET[ UV[OI VTE UJIT M TIEO VMPI YMVL EHESVMXI M UVE GI RTPEOM EVMR 3 3.. C D .R SOIH 1I ITEVMXI ,HXITUETMEO 8IVYRTN].R1,8 7 A 6M E H IO .R SOIH KI ITEVMXI EHXITUETMEO IVYRTNU 3 8I T3: C D :MZIO , 5 -R UPEOMU 8 MO ITPE RLE 0TLE E H 5TMUL E U SITXMUIH SMZIO OIXIO HRPEM EHESVEVMR YMVL KI ITEVMXI EHXITUETMEO IVYRTNU 3 . : C D .[., , 4 2RJJPE 0 I K :ETN 4 A BL : 3UROE 5 EI NR , 0JTRU E H ETTIOO .[GEHE .[GOI GR UMUVI V EHXITUETMEO HRPEM EHESVEVMR 3 3.76 C D , , 5 EMVR A ULMN E H 2ETEHE ,U[PPIVTMG VTM VTEM M K JRT U SITXMUIH HRPEM EHESVEVMR 3 3.76 45