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

社内輪講資料 / semi-supervised learning

社内輪講資料 / semi-supervised learning

2020.05.22 社内勉強会にて発表に使用した資料です。

Takumi Karasawa

May 22, 2020
Tweet

More Decks by Takumi Karasawa

Other Decks in Research

Transcript

  1. 8 8 C . 68 n - 8 A 8

    te g]K [ S c N ] i M t]m L h oIrn l m[c [v m e l [ a KbT n S NR P aS n oIrn KM n u[T dy ]“ K KbT b[ M N] KbT n u A C 8 A 8 8 6 sIp (,) n J9 0 A 2 0 1DD 1 1 n 7 J / : 7 J 8 EH : , . EH , 8 EH : , n :A D A ) 5DJ : , 49 3 9 M , L(( 8E  2
  2. .31 5 3 2 3 2 1  -3 3

    32 35 C 4 3 2 1 1 35 , 3
  3. - . - , - - -, - . -

    . - -, a - -, a ea r - -, a S - r - -, - . - -, - . n - -, - . u S ilp U n ( - -, - . n y o T - - -, - .T - -, a - -,a S Svs Cbc dg t ➤ - -, a - -, a M Tm C C LT - -, a h u C m C vs S - - - 5
  4. ,205 2 ( ) 28 2 2 2. M L

    C n 28 2 2 2. M M gh 0 . 0. n g u 1 1 :1 5 n dbMTf 28 MiSa ZC-5 e n 4 5 5, 2, 0 -, ky sb n 1 - -: - , - 15 al T o n cMdb n m e ch id n p f n S ky n t r - - - 6
  5. ) - 0.26 0 ( 6 ,-0 0 ebl C

    ,-0 0 e v “ ghc i d bys p m e Lxu n bo ➤ 6 ,-0 0 e n b| aCt n L| a 0 0 0 0, 6 6 , 6 2, 0 0 0 , - nr n T MS 0. 6 , () 7
  6. 0 , 1 2 2 T S m iln /

    / drv u p g M S T y 8 2 T a sc C n /0 2 /0 2 m m n /0 2 m b y O n 8 2 y b yO [ 2 t hvd T S L ] ➤ - / .C y [ edovd b 8
  7. . , .,. . 02 0 002 - 0 0

    -- 00 - 0 - 20 9
  8. )., . - : . .- ML C M C

    S T n (. : . .- . : n - n - n - - 10
  9. 2 . . - 2 2,. 2 g dL L

    T C e L M L n . . , a i b . . ci b h 11
  10. , , , n b l a d g p

    iML o p iM n C . . , n . . , m C h T ec n b o p i C - 12
  11. . , . m so . p e e b

    L . c n . l a m L n d m . . n . . d m y d . Cg y T M i tfh L 13
  12. . 6 , / 6 ) 0 v y ed

    - : /6 - : ) -: 66 -6 . d n /6 - : ( shoi d C shoi d T b T M - : fs t T hoi u b 6- - : 6- / / : r n ( w C 6- n 6 0 ( w C 6- n - 2 60( : C 6- 6-. 66 0 6-. 66 0 C f f Thoi e : 0 bac lmgCnpi d L d Ld 15
  13. . . ,. U n n n . CM L

    S n T . - - - 1st distinction: Inductive / Transductive
  14. . . - ,. . . M b L lg

    e aI n . n h TL do c i C lg n . n n L m C lg n a i n d n lb e d c e lb / ( /( /( /()( /( ( /(1 / /(1 18 Setting: Inductive methods: Transductive methods: for
  15. - . - , , - - , g u

    - -, We d b hs - -, M T L IC r C UT n . - , lcad n - -, - - . n L n - - -, - , tpyuio m vU 2 19 2nd distinction for inductive: how to incorporate unlabelled data
  16. . - 20 20 - . . . ,, -

    , . - . M d f eL b ah cg M G L M T C - (- 2 ) 2nd distinction for transductive: which stage of learning process
  17. . , .,. . 0 . . 3. 0 -

    4 : . 1 .- . . 14 4 14 1 22 .31 : . 1 .- 3. 0 - 21
  18. . . - ,. oL T . i c .

    . M g d s g d v il C L cfeh n s oL T . . . g d v i ma g d s n v i ma . , s . . M g d il . y g d t M il v il M g d n bmhl r u T . p n n i c Mo . n Ti c 22
  19. 2 : -5 : 5 . 1 5 5 c

    gb - 5 5 1: 0 - 5 1 5 5 [ n L T[ :1 5 ap n I t[sn h S dbf[ :1 5 a L n Ih S dbf[sn :5 C 5 : C 5 Ie d[ P W w y T[mM TL u n ,5: 1 ] T[ :1 5 [ v 5 5 l L o r n 1 5: 1 a t[ :1 5 i n 5 :125::5 2 5 23
  20. 1 , 2 0 s]h bea n Y R C

    D 2 0 5 k om n .-0 0 su p j M gi d T n . 1 0 19 2 2 n [L n C 01 f c wy 2 2 5 2 l M v n C 01 f c rt n 2 0 ). 1 1 - -( 1 - -( 24
  21. , 1 : 2 : 3 0 C 3 :01

    :: 0 02 n 2 0: d bL n l yd bg iM n evhu s d N H wki “ n lrf o 3 :01 :: bT b n -- I . 3 :01 : : 0 3 2 C 3 : 0 3 3 0: D , D n h l d i - a g n o e b ip n D 0 3 ptmfsPa cb ). 1 1 - -( 1 - -( 25
  22. 1 . 1 - 1, - i 1 , b

    n t 1 , i C n - jrw o n 1 - y C ➤ 1 n -1 - e 1 f n L ni 1 e i fc e n - 1 1 , i a h , f 1 i f z n gTg gd i f n f n flumusv M f n - - . 1 - 26
  23. ., . n d C i Sg cABg L M

    c n E . . Cag T e : . . hb B n n . : . . . . . n . . . 27
  24. , . - , , , . , , ,

    , . F , , M , , ML P T n , , , n ( ( n ( ) n , ., , , - C , - n ( ( S CPe ( h gfCPA a P n b c P d A ) DP n , - n n ( ) ( ) 28
  25. 2 - 2 - , s T - - -,

    t i g - -, - 2- - o LT n v lx - -, d cabh L 2 -22-, 2- w LNnrGy n - u N - - -, 2- M C e fxLNn T n - , ➤ 2 ,- n - -, - , ➤ - mp n . 2, ➤ . 2, n - - - ,-2 - . 29
  26. . . - ,. C uie h m dL lC

    ➤ . . M fzvT . . gsTw a b x t c y r n n -. n 3 ( ( )3- ) a Su V ( : a : : : bieM g : a Su n ( S s d p N r n ( Sl v P : n , . n . , . . opd n . , ,. . . . n a : : Pm ieS : : . - - - - 3 30
  27. , . - , n 2 . , 2 T

    lgn o ch n sCN i d N T M ko n g e i d am p g e i d g e N n L g e i d ,2 , , N n - - .2 bpkn y s n , , g e r N N ,2 , , T n u v u ko t N - . . . - . 3 - - . 31
  28. (., . - C n --. . n l g

    c ) n ) v w ) ( ) ) , ) ) c ) ) ( a n ) k kg dr ) n . - . . . n n u z o u n u ) ) f ) ) n A -. n . . . . n . . , . n ) - . ) ( n u s i tx p k u em N ) ) n u p n . . .- - . . . - . 3 - - . 32
  29. ( . (, n , , C, E mdM i

    okM . , . c G N gM V T r n , , A , EnbkM NhaL , e pk f l , n , , A n , , , , ) n , , , , . - . 33
  30. , ) - . n . n . - n

    - n . , . . . n . - - U MM C n ( L T C , . . . . - n . - - . P CR n - 2 - 35
  31. (- - , - - . - ,- , n

    n L e fT n - - -, i d lnfa - -, k fmcn t L - -, h gau b v a bT n Po p sA n - - -, - - - -, C L T y M rw o A n 36