Upgrade to Pro — share decks privately, control downloads, hide ads and more …

寝ながら学べるDeep Learning

ueniki
December 03, 2019

寝ながら学べるDeep Learning

Deep Learningの数学的な基礎を簡単に説明します。
急速に普及をみせるDeep Learnig・AI技術ですが、多くの人は理解せずに語っているのが現状です。
高校数学程度の簡単な知識を使って、Deep Learnigのコアを解説します。

# AI # 機械学習 # Deep Learning # ディープ・ラーニング

ueniki

December 03, 2019
Tweet

More Decks by ueniki

Other Decks in Technology

Transcript

  1.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  2.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  3. 2015 年 12 月 2 日 株式会社野村総合研究所 日本の労働人口の 49%が人工知能やロボット等で代替可能に ~

    601 種の職業ごとに、コンピューター技術による代替確率を試算 ~ 株式会社野村総合研究所(本社:東京都千代田区、代表取締役会長兼社長:嶋本 正、以下 「NRI」 ) は、英オックスフォード大学のマイケル A. オズボーン准教授およびカール・ベネ ディクト・フレイ博士*1 との共同研究により、国内 601 種類の職業*2 について、それぞれ人 工知能やロボット等で代替される確率を試算しました。この結果、10~20 年後に、日本の 労働人口の約 49%が就いている職業において、それらに代替することが可能との推計結果 が得られています。 この共同研究は、NRI 未来創発センターが「“2030 年”から日本を考える、“今”から 2030 年の日本に備える。」をテーマに行っている研究活動のひとつです。人口減少に伴い、 労働力の減少が予測される日本において、人工知能やロボット等を活用して労働力を補完 した場合の社会的影響に関する研究をしています。 ▪ 日本の労働人口の約 49%が、技術的には人工知能等で代替可能に
  4.   G MM C N I BLL , -,

    ) -, ) , % % % 6: % % % ( % ) 2 GDB % H I % 2 GDB 45: 90 V XS]V [ 3B BN G 1BGP ID 1BB IN 9B N C GBLM. : L MM ID 3 H I 5BPBG 8BLC LH I B I 4H DB BN 0G MM C N IR L P %
  5.  , : d b r x ty e c

    kj wn ud mw h a kj i d d” e xso S1: 1: T d a fj i d” c vzpd i vzp d l g d e h a b A E EA I H 7 GA A 9 : A7 E E 7 : E GA A A 9 " . E I E A EG E 7 0 E M1: 1: N - GE 9 , 7" E I E A EG E 7 7 A7 9 A EE 9 C 7" 1 A E C A 7 I C 7 I E E A G7 " A E GA A 9 9 H 7 I EA I I E E 7 : E A 9 " EG 9 I E E 7 1: 1: A 0 E GE 9 , 7" 1 C 9 I E 7 7 A7 9 A 7 7 7 E 9 C 7" 1 A E H C A 7 I C 7 I A 7 G7 " A E 0 - A:I K" M 2A IE 9 A N" “ 1 IEC E"
  6. 7 a A 9 A 9 7 1H 0 ,9

    9 1 :1A K ” Noxuz vyb hg "tk ra jtxn e S hg SfM a a b upl .71: .719 a cg SfM a swma fMswm a i d S aTb e SM N,9 91 :1A 9 1 - 19 " 97 1 9 19 2 97 19 9 A9 1 - 9 9 1 .71: .719 K A 9 A 9 A9 1 A 1A . 1 G 19 1 1A 1 9 1 1 9 N,9 9 1 :1A 9 1 19 " E A 9 9 9 19 2 97 9 A9 1 0 9 9 1 .71: .719 9 1 19 . 1A A 1A A9 1 A 1 2 9 1A . 1 E A G 19 1 1A 1 1 1 1 9 “ . 9
  7.   (E IG C L : A :: /

    / / : :: / )( . : / : / / : / :
  8.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  9.  • e P c l k i ad e

    U • 0 .- , 1 T -.- / 1 h 0
  10.  • T • S T 3 . -32 P

    . . 2 U • • U • .2 • 3 D B • .
  11. 1) 4 23( 4      

           
  12.          

                 
  13.          

                                   
  14. 1         

      
  15.          

         4 2 1 3   
  16.          

                                        1 2
  17.          

        1 2      
  18.          

                                      1
  19.          

          1    
  20.          

                 
  21.  2133 - 4-10 V 1 2133 5 - 4-10

    C 1 2133 5 - 4-10 K K K K
  22.      -(, )%03 .  !&

    $ '+ #!& 1* '+ 2!&  /"
  23.  *()3'&!%F #"-8<6B2A'' #" 0>:5   H  @9

     +1''4/E  ''$ ,2ACD 7 . '&!% = ?;G *()3 =
  24.  ( ) 8 8 8 1 0 ! =

    #$ #% ⋮ #' 0 8 8 28 8 7462 8 5 936 0
  25. L D

  26.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  27.   a p nD 1 1 ( ) -

    - ei D 2 2 2 2 2 2 !" !# g D 5 9 7D 5 r - L - 9
  28.  : "# $# + "& $& + ' =

    0                 ("# $# + "& $& + ' ≥ 0) ("# $# + "& $& + ' < 0) ($## , $&# , 0_# ) $# $& =1
  29. : "# $# + "& $& + ' = 0

    1 1 1 ("# $# + "& $& + ' ≥ 0) ("# $# + "& $& + ' < 0) −1 ($## , $&# , 2_# ) $# $& −1 0 =1
  30. : "# $# + "& $& + ' = 0

    1 1 1 ("# $# + "& $& + ' ≥ 0) ("# $# + "& $& + ' < 0) −1 $# $& 0 = 1 1 "# $# + "& $& + ' ≥ 0 −1 "# $# + "& $& + ' < 0 , 1 "# "& ' ($## , $&# , 0_# ) =1
  31. ! "# "$ %# %$     

    $#    !%&   "
  32. ! "# "$ %# %$     "$,+

    !   (#)-.&   '*  ! = ' 1 "# %# + "$ %$ ≥ + −1 "# %# + "$ %$ < + −1 0 %
  33. ! "# "$ %# %$     

    ! = ' 1 "# %# + "$ %$ ≥ + −1 "# %# + "$ %$ < + 1
  34. ! "# "$ %# %$     !"#,098

    +   4'/*5:;2 $  36  ! = ' 1 "# %# + "$ %$ + * ≥ 0 −1 "# %# + "$ %$ + * < 0  - (-1 5:2.:& 7%  )
  35. ! " #$ 1 &$     &#

      ! = ( 1 #) &) + #$ &$ + " ≥ 0 −1 #) &) + #$ &$ + " < 0  $ !$' )+( %+   * " &) #)
  36. ! " #$ %$      

       %& #& ! = ( 1 * + , ≥ 0 −1 * + , < 0 * + , = " #& #$ 1 %& %$ 1
  37. ! " #$ %$   4 2 0 1

      %& #& ! = ℎ ) * + ) * + = " #& #$ 1 %& %$ ℎ % = - 1 % ≥ 0 −1 % < 0   . 0 1
  38. ! "# $#   2 . 04 1 

     $% "% & = ℎ ) * + ) * + = ! "% "# 1 $% $# ℎ $ = - 1 $ ≥ 0 −1 $ < 0   ℎ( ) &   4 04 1
  39. ! " # = %& %' (1 (_ ⋅ !

    " # + ,& ,' 1 %& %' ! " # ≥ 0 ! " # < 0 : ! " # = 0 ( = 1 ( = −1 (%&& , %'& , (_& )  =1 (_ ⋅ ! " # ≥ 0 (_ ⋅ ! " # < 0 (1 (_ ⋅ ! " # ≥ 0 ! ) (
  40. ! " # = ) ) %& %' 1 )

    ( ( )& )' 1 %& %' ! " # ≥ 0 ! " # < 0 ) : ! " # = 0 / = 1 / = −1 (%&& , %'& , /_& )  ) =1 6 ! = 7 8 9(/_ ⋅ ! " # ) 9 % = ; 0 (% ≥ 0) −% (% < 0) 6 ! = 0 ( (!
  41. ! " # = %& %' ( = ℎ !

    " # = * 1 (! " # ≥ 0) −1 (! " # < 0) 1 ! 1 1 2 3& 3' 1 %& %' : ! " # = 0 (%&& , %'& , (_& )   =1 (_ ⋅ ! " # ≥ 0 (_ ⋅ ! " # > 0
  42. … "# "$ : & ' ( = 0 &

    ' ( = 1 + ,# ,$ 1 "# "$ . = 1 . = −1 & ' ( ≥ 0 ("## , "$# , ._# ) & ' ( < 0
  43. … "# "$ : & ' ( = 0 &

    ' ( = + ,# ,$ 1 "# "$ . = 1 . = −1 & ' ( ≥ 0 1 ……… ("## , "$# , ._# ) & ' ( < 0
  44. 1 1 … "# "$ 1 : & ' (

    = 0 & ' ( = 1 1 + ,# ,$ 1 "# "$ ………  1 . & = / 0 1(3_ ⋅ & ' ( ) 1 " = 7 0 (" ≥ 0) −" (" < 0) ("## , "$# , 3_# ) 3 = 1 3 = −1 & ' ( ≥ 0 & ' ( < 0
  45. !" !# : % & ' = 0 % &

    ' = * +" +# 1 !" !# (!"" , !#" , 0_" ) 0 = 1 0 = −1 % & ' ≥ 0 % & ' < 0 (( 1 ) 1 ( 1
  46.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  47.      !" !# $ 0 0

    0 1 0 0 0 1 0 1 1 1 
  48. ! " #$ %$ 0 1 2 5  

    %& #& ! = ℎ ) * + ) * + = " #& #$ 1 %& %$ ℎ % = - 1 % ≥ 0 −0 % < 0  . 14      1
  49.   !" !# $ 0 0 0 1 0

    0 0 1 0 1 1 1 !" !# $ 0 0 1 1 1 0 1 1 1 1 0 0     
  50.   !" !# $ 0 0 0 1 0

    1 0 1 1 1 1 1 !" !# $ 0 0 0 1 1 0 1 1 1 1 0 0     
  51.   !" !# $ 0 0 0 1 0

    0 0 1 0 1 1 1 !" !# $ 0 0 1 1 1 0 1 1 1 1 0 0     
  52.   !" !# $ 0 0 0 1 0

    1 0 1 1 1 1 1 !" !# $ 0 0 0 1 1 0 1 1 1 1 0 0       
  53. !" !#      $  

    %# %"    
  54.  !" !# 0 0 1 0 0 1 1

    1 &" &# ' 1 0 0 1 1 1 1 1 0 1 0  1       
  55.      !" #$ #"  

    !$ %     
  56.  !" !# $" ℎ( ) ( ℎ( ) )"

    ℎ( ) $# )# )* ( = , -" #$" + -# #$# + /# + 0 = , -" # , -"" " !" + -#" " !# + /" " + 0 +-# # , -"# " !" + -## " !# + /# " + 0 +/# + 0 = ⋯ -" # -" # -#" # -## # -"" # -"# # ℎ = ,! + 0 ( !" , !# 1
  57.  !" !# "  $" " ℎ( ) (

    # # ℎ( ) # # )" ℎ( ) # # $# )# )* +" # +" # +#" # +## # +"" # +"# #        !
  58.  !" !#   $"  ℎ( ) (

      ℎ( )   )" ℎ( )   $# )# )* +" # +" # +#" # +## # +"" # +"# #        
  59.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  60.     !" !# $" $# $% 1

    2 3 4 5 6 , 1 2 3 4 5 6 !" !# = $" $# $% . /
  61.  1 2 3 4 5 6 1 2 3

    4 5 6 = 9 12 15 19 26 33 29 40 51 * + , 1×1 + 2×4 5×2 + 6×5 / 0 1 3 × 4 4 × 6 3 × 6
  62. !" !# 1  %" " ℎ( )  )"

    " *"# " *"" " +" " = +" " *"" " *"# " +# " *#" " *## " +- " *-" " *-" " 1 !" !# ./ 0 12 = %" " %# " %- " = 3/ 4/ = ℎ 3/ ℎ( )   %# " )# " ℎ( )   %- " )- " = ℎ %" " %# " %- " = )" " )# " )- "
  63. 1 ℎ( )   %& & '& & ℎ(

    )   %( & '( & ℎ( )   %) & ') & ℎ( )   %& ( '& ( ℎ( )   %( ( '( ( ℎ( )   %) ( ') ( *+ , -. = 0& ( 1&& ( 1&( ( 1&) ( 0( ( 1(& ( 1(( ( 1() ( 0) ( 1)& ( 1)( ( 1)) ( 1 '& & '( & ') & = %& ( %( ( %) ( = 2+ 1&) ( 1&( ( 0& ( -+ = ℎ 2+ 1&& ( = ℎ %& ( %( ( %) ( = '& ( '( ( ') (
  64. 1 g( ) %& ' (& g( ) %) '

    () ℎ( ) %& ) +& ) ℎ( ) %) ) +) ) ℎ( ) %' ) +' )       ,&' ' ,&) ' -& ' ,&& '
  65. !" !# ℎ( )   '" # (" #

    ℎ( )   '# # (# # ℎ( )   '# # () #  '" " ℎ( )  (" " ℎ( )   '# " (# " ℎ( )   ') " () " g( )   '" ) +" g( )   '# ) +# 1 1 1
  66.          

        ℎ " = $ 1 " ≥ 0 −1 " < 0
  67.      01 ℎ " = 1

    1 + &'( .01 ℎ " = ) 1 " ≥ 0 −1 " < 0
  68.      ℎ " = $ 1

    " ≥ 0 −1 " < 0 . 0 0 ℎ " = $ " " ≥ 0 0 " < 0
  69.      ℎ " = $ 1

    " ≥ 0 −1 " < 0 10356 56 ℎ " = $ " " ≥ 0 0 " < 0 . 4 256 ℎ " = 1 1 + +,-
  70.    , L r eh Ll Y L

    19 6 uhL 55 1 85 k Lo w Lc × L a c 02 551 I T M NIL Sn c L C
  71.     g( )   $% &%

    g( )   $' &' g( )   $( &( ⋮ *+,- ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮    *( */   
  72.     g( ) # # $% &%

    g( ) # # $' &' !  g( ) # # $( &( ⋮ * $+ = -./ $+ ∑ 12% 3 -./ $+ & = * $+ !   "  
  73.     g( ) $% &% g( )

    $' &' g( ) $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮     
  74. g( ) $% &% g( ) $' &' g( )

    $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮      0 = *( *. ⋮ *+,- = 0.05 0.7 ⋮ 0.1
  75.    # "&(, %8> 0!2 .= (" #

    ⋯ "% ⋯ ) "' = 0 ⋯ 1 ⋯ 0 7; 0(,!2 +9<)  '1-$*64/   :50!2 +9< 3   
  76.  264( >[VU!SNT L @;?HOB?Q :XZ @;?H?H*(?H"8%7'3 FP=R  -568.)+9$

    264(<MI1#"& 0#2264(MJ <M/,L?H[ 0#2264( ?H!DEKGC   264(  S@;?H264(Y>AW 
  77. ( )) ( ! = − $ % &% log

    *% = *+ *, ⋮ *. 6 6 = &+ &, ⋮ &% K 8 0 - .04 K 2033 - 42015 ( ) / * = −012/
  78.   h L L 5 h L h D

    P ea o L5 h 5   r h t n y 0 1. . 2, . g p 6 i p 0
  79. g( ) $% &% g( ) $' &' g( )

    $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮     61 7 6 0 6 6 . 0= 0= 5 6
  80. g( ) $% &% g( ) $' &' g( )

    $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮     0 .0 6 7 = . 01 6 6 0 =5
  81.  ! " " # ! !$ # ! !$%&

    : ! !$%& = !$ − ) * +# +! ,-,. *
  82.    !" !# = !" !%&& !" !%'&

    !" !%(& !" !%&' !" !%'' !" !%(' #)*& = #) − , - !" !# #.#/ " # = %&& %'& %(& %&' %'' %(' 0 ( (  ) ( (
  83.  g( ) $% &% g( ) $' &' g(

    ) $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮     
  84. g( ) $% &% g( ) $' &' g( )

    $( &( ⋮ *+,- *( *. ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ … ⋮ ⋮ 0 = *( ( *( . ⋮ *( +,- *. ( *. . ⋮ *. +,- ⋱ *3 ( *3 . ⋮ *3 +,- = *( % *( ( ⋮ *( ' *. % *. % ⋮ *. ' ⋱ *3 % *3 ( ⋮ *3 ' 4 = − 6 3 6 7 89: log &9: ) ( = 8( % 8( ( ⋮ 8( ' 8. % 8. ( ⋮ 8. ' ⋱ 83 % 83 ( ⋮ 83 ' = 1 0 ⋮ 0 0 1 ⋮ 0 ⋱ 0 0 ⋮ 1 = 0.05 0.1 ⋮ 0.2 0.03 0.4 ⋮ 0.15 ⋱ 0.35 0.1 ⋮ 0.01 0 D .1
  85.  r l n D ip ea 4 202 r

    l ea N A h ea o r l LI R N I r l 4 202 gG C 4 202 r l 14 3
  86.  PL C C R a C CR a C

    U R a PL ⋯ - 1
  87.  ⋯ "# "$ "%& "%' "() ⋯ "*&+ ,

    = ./ ⋮ "$ ⋮ "%& "%' ⋮ "$1%& ⋮ .234 3 3 3 3 "$1%& ⋯ ⋯ ⋯ ⋱
  88.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 8 ⊛ 2 1 1 0 2 0 1 0 2 1
  89.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 8 8 17 ⊛ ⊛ 1
  90.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 8 8 17 8 17 17 ⊛ ⊛ ⊛ 1
  91.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 2 1 1 0 2 0 1 0 2 8 8 17 8 17 17 8 17 17 7 ⊛ ⊛ ⊛ ⊛ 1 1
  92.   2 1 1 1 2 0 3 1

    0 1 2 1 3 3 2 0 2 1 1 0 2 0 1 0 2 ⊛ 8 17 17 7 2 10 19 19 9 +  02&    "(  -%'. ) /  13 )   *3 "(&  #,!$+ℎ( )02 -( )
  93.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 ⊛ 2 1 1 0 2 0 1 0 2 7 6 7 10 11 3 12 4 8 5 9 8 15 11 17 7 0 0
  94.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 0 0 3 2 2 1 1 3 1 2 1 3 0 1 0 1 3 1 1 2 ⊛ 2 1 1 0 2 0 1 0 2 15      
  95.  2 1 1 1 2 0 3 1 0

    1 2 1 3 3 2 0 0 0 3 2 2 1 1 3 1 2 1 3 0 1 0 1 3 1 1 2 ⊛ 2 1 1 0 2 0 1 0 2 ⊛ 2 1 1 0 2 0 1 0 2 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 0 0 3 2 2 1 1 3 1 2 1 3 0 1 0 1 3 1 1 2 15 15 13 
  96.    !" = " + 2& − ("

    ) + 1 !+ = + + 2& − (+ ) + 1 ((", (+)  (", +)  (!", !+)  (!", !+)  ((", (+)S   S
  97. 40 ⊛ 2 0 2 1 3 2 3 1

    2 2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 1 2 3 1 1 3 1 1 2 1 2 1 1 0 2 0 1 0 2       R B 3 B 3 G '(, '*, + ,(, ,*, + '(, 0*, 1
  98. 40 ⊛ 2 0 2 1 3 2 3 1

    2 2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 1 2 3 1 1 3 1 1 2 1 2 1 1 0 2 0 1 0 2 55       B 3 B 3 G R (), (+, , -), -+, , (), 0+, 1
  99.       40 ⊛ 2 0

    2 1 3 2 3 1 2 2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 1 2 3 1 1 3 1 1 2 1 2 1 1 0 2 0 1 0 2 55 61 B 3 B 3 G R )*, ),, - .*, .,, - )*, 0,, 1
  100. 40 ⊛ 2 0 2 1 3 2 3 1

    2 2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 1 2 3 1 1 3 1 1 2 1 2 1 1 0 2 0 1 0 2 55 36 61       B 3 B 3 G R )*, ),, - .*, .,, - )*, 0,, 1
  101. ! " # 1 %" ! &" &# 1 1

    1 %#  ⊛
  102. ! " # 1 %" %# ⋮ ! '" '#

    1 1 1  ⊛
  103. ! " # ! $" $# ⋮ &" &# $'

    1 $' 1 1 ⋮  ⊛
  104. ! " # ! $" $# ⊛ ⋮ '" '#

    $( 11 1 ⋮ 
  105. ! " # ! $" $# ⊛ ⋮ + 1

    1 $)  $) *" *# *# *" $)        +( )    *#, *", ! *#, *", !, $) *#, *", ! *#, *", !
  106. 2 1 1 1 2 0 3 1 0 1

    2 1 3 3 2 0 2      
  107. 2 1 1 1 2 0 3 1 0 1

    2 1 3 3 2 0 2 2 3      
  108. 2 1 1 1 2 0 3 1 0 1

    2 1 3 3 2 0 2 2 3 2 3 3      
  109. 2 1 1 1 2 0 3 1 0 1

    2 1 3 3 2 0 2  2 3 3 3 2 3 2 3 3     
  110. 2 1 1 1 2 0 3 1 0 1

    2 1 3 3 2 0 2  2 3 3 3 2 3 2 3 3 M M X A
  111. 3 2 0 2 1 3 2 3 1 2

    2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0         %&, %(, )  3 %&, 0(, )  2
  112. 3 3 2 0 2 1 3 2 3 1

    2 2 1 2 3 0         %&, %(, )  %&, 0(, )  2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0 3 2 2 3
  113. 2 0 2 1 3 2 3 1 2 2

    1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0         %&, %(, )  3 3 3 2 2 3 3 %&, 0(, ) 
  114. 3 3 2 2 0 2 1 3 2 3

    1 2 2 1 2 3 0 2 1 1 1 2 0 3 1 0 1 2 1 3 3 2 0       1 %&, %(, 1 1 )&, )(, * )&, 0(, * 3 2 3 2 3 3 3
  115.     / // 256×256×3 256×256×64 128×128×128 128×128×128

    64×64×256 32×32×512 16×16×512 e dN R oC h P: c an d . / f C c o f h R
  116.  G MM C N I BLL , -, )

    -, ) , % % % 6: 9BM BN % % % ( % ) 2 GDB 2 5B BN % H I % 2 GDB 0 VX 45: 90 [S ] 3B BN G 1BGP ID 1BB IN 9B N C GBLM. : L MM ID 3 H I 5BPBG 8BLC LH I B I 4H DB BN 0G MM C N IR L P %
  117. : : : 54 G  1 , 01/ G

    3×3 uc n V a R IC g Shd a VP I a P 1 R i S L I C 1 w so a S GG 55 52 6 . : S pc y r g eyd D t
  118.  , v ,01 L 0 5E 2 9 D

    C . 4 5 9 4 4 4 C D 5 4 G ith softmax loss as the classifier (pre- 1000 classes as the main classifier, but ence time). of the resulting network is depicted in odology rks were trained using the DistBe- achine learning system using mod- and data-parallelism. Although we plementation only, a rough estimate ogLeNet network could be trained to w high-end GPUs within a week, the the memory usage. Our training used tic gradient descent with 0.9 momen- ng rate schedule (decreasing the learn- 8 epochs). Polyak averaging [13] was l model used at inference time. methods have changed substantially ing to the competition, and already re trained on with other options, some- with changed hyperparameters, such arning rate. Therefore, it is hard to ance to the most effective single way s. To complicate matters further, some ainly trained on smaller relative crops, , inspired by [8]. Still, one prescrip- to work very well after the competi- ng of various sized patches of the im- ributed evenly between 8% and 100% h aspect ratio constrained to the inter- found that the photometric distortions ] were useful to combat overfitting to s of training data. 14 Classification Challenge sults 4 classification challenge involves the image into one of 1000 leaf-node cat- et hierarchy. There are about 1.2 mil- ng, 50,000 for validation and 100,000 Each image is associated with one , and performance is measured based g classifier predictions. Two num- rted: the top-1 accuracy rate, which truth against the first predicted class, te, which compares the ground truth edicted classes: an image is deemed the ground truth is among the top-5, in them. The challenge uses the top-5 input Conv 7x7+2(S) MaxPool 3x3+2(S) LocalRespNorm Conv 1x1+1(V) Conv 3x3+1(S) LocalRespNorm MaxPool 3x3+2(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) MaxPool 3x3+2(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) AveragePool 5x5+3(V) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) AveragePool 5x5+3(V) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) MaxPool 3x3+2(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) AveragePool 7x7+1(V) FC Conv 1x1+1(S) FC FC SoftmaxActivation softmax0 Conv 1x1+1(S) FC FC SoftmaxActivation softmax1 SoftmaxActivation softmax2 Figure 3: GoogLeNet network with all the bells and whistles. 4 5C L , mWn rS 4 5C ot uVR I 1×1 f hc s pN PV dgieh mWa l I (a) Inception module, na¨ ıve version 1x1 convolutions 3x3 convolutions 5x5 convolutions Filter concatenation Previous layer 3x3 max pooling 1x1 convolutions 1x1 convolutions 1x1 convolutions (b) Inception module with dimensionality reduction Figure 2: Inception module efficiency during training), it seemed beneficial to start us- ing Inception modules only at higher layers while keeping the lower layers in traditional convolutional fashion. This is not strictly necessary, simply reflecting some infrastructural inefficiencies in our current implementation. A useful aspect of this architecture is that it allows for increasing the number of units at each stage significantly without an uncontrolled blow-up in computational com- plexity at later stages. This is achieved by the ubiquitous use of dimensionality reduction prior to expensive convolu- tions with larger patch sizes. Furthermore, the design fol- lows the practical intuition that visual information should be processed at various scales and then aggregated so that the next stage can abstract features from the different scales simultaneously. The improved use of computational resources allows for increasing both the width of each stage as well as the num- ber of stages without getting into computational difficulties. One can utilize the Inception architecture to create slightly inferior, but computationally cheaper versions of it. We have found that all the available knobs and levers allow for a controlled balancing of computational resources resulting in networks that are 3 10⇥ faster than similarly perform- ing networks with non-Inception architecture, however this requires careful manual design at this point.
  119.  , I ,01 0 5E 2 9 D C

    . 4 5 9 4 4 4 C D 5 4 G input Conv 7x7+2(S) MaxPool 3x3+2(S) LocalRespNorm Conv 1x1+1(V) Conv 3x3+1(S) LocalRespNorm MaxPool 3x3+2(S) Conv 1x1+1(S) Conv 1x1+1(S) Conv 1x1+1(S) MaxPool 3x3+1(S) DepthConcat Conv 3x3+1(S) Conv 5x5+1(S) Conv 1x1+1(S) Figure 3: GoogLeNet network with all the bells and whistles. type patch size/ stride output size depth #1⇥1 #3⇥3 reduce #3⇥3 #5⇥5 reduce #5⇥5 pool proj params ops convolution 7⇥7/2 112⇥112⇥64 1 2.7K 34M max pool 3⇥3/2 56⇥56⇥64 0 convolution 3⇥3/1 56⇥56⇥192 2 64 192 112K 360M max pool 3⇥3/2 28⇥28⇥192 0 inception (3a) 28⇥28⇥256 2 64 96 128 16 32 32 159K 128M inception (3b) 28⇥28⇥480 2 128 128 192 32 96 64 380K 304M max pool 3⇥3/2 14⇥14⇥480 0 inception (4a) 14⇥14⇥512 2 192 96 208 16 48 64 364K 73M inception (4b) 14⇥14⇥512 2 160 112 224 24 64 64 437K 88M ⋮ ⋯ 256 = 64 + 128 + 32 + 32 NL P
  120.    h i t woi t c egr

    d np f m ʼ b k , 4 :- 6 0. /97 4 :- 1/9385 598 1. 2/ /97
  121.     "    !", !$,

    !%, ⋯ , !'  !   
  122.    ( ) !", !$, !%, ⋯ ,

    !' ( (", ($, (%, ⋯ , () *) ()+" ( )
  123.    L u r Ln !", !$, !%,

    ⋯ , !' Y U i d ) ye o . ( . (" ($ (% () (* (+ (, -" -$ -% -) -* -+ . ( . H kLyeU v U a (", ($, (%, ⋯ , (. -. (./" U v b yeL w F Usl m yeL vU t H g H
  124.      ( N R ) !

    " = 1 " = 2 " = 3 ⋯
  125.  2 2 ! 2 ! !"#$ !%& ⋯ ($)*

    +$)* ($ +$ ($,* +$,* 2 - − 1 - ℎ$)* ℎ$ ℎ$,*
  126.  a 2 ! ! !"#$ !%& ⋯ t m)$

    f m *$+, )$+, *$ )$ *$-, )$-, m a . − 1 . ℎ$+, ℎ$ ℎ$-,
  127. 552 . 2 5 -5 -1 00 5- 0 :

    / / / / 1 !"" !#$% !&' 1 0 0 0 0 0 0 1 0 0 1 0 0 0 1 0 0.3 0.1 0.0 0.6 0.9 −0.1 0.3 0.5 0.1 1.0 −0.9 −0.5 0.3 0.7 −0.1 0.3 0.05 0.08 0.8 0.07 0.08 0.1 0.7 1.2 0.01 0.95 0.02 0.02
  128.      !" ℎ"$% &" ℎ" '()ℎ

    '()ℎ !"*% +,, - ℎ"$% + /00 + '()ℎ + +12 - !" + /34 +567 - ℎ" + /89"
  129. ) ( ( !" !"#$ ℎ"#$ × + × ()*ℎ

    × ()*ℎ + + + ℎ" " ! ," -" "     !" " ℎ" #$!
  130. ( ) !" !"#$ ℎ"#$ × + × ()*ℎ ×

    ()*ℎ + + + ℎ"   ," -" !"    
  131.    !" !"#$ ℎ"#$ × + × ()*ℎ

    × ()*ℎ + + + ℎ" " ! ," -" ." -" = + 012 3 4 ," + 055 3 4 ℎ"#$ + 63 $*+ #    %(  "'&)
  132.    !" !"#$ ℎ"#$ × + × ()*ℎ

    × ()*ℎ + + + ℎ" +" )" ," ," = + ./0 1 2 3" + .44 1 2 ℎ"#$ + 51 3" 6" ,.+"#!1( '* /6%7  &-  04  +"#!3$ 25
  133.    !" !"#$ ℎ"#$ × + × ()*ℎ

    × ()*ℎ + + + ℎ" 5) 4) ," ," = + ./0 1 2 3" + .44 1 2 ℎ"#$ + 51 3" 6" 4)*#:/ '3$(28% 5)*#<+ '3$(28%  .  1&0 7?,@! -6" 9=  !5)*#<+ ;>
  134.       !" !"#$ ℎ"#$ ×

    + × ()*ℎ × ()*ℎ + + + ℎ"   ," -" !" ~ -" = + 012 3 4 5" + 066 3 4 ℎ"#$ + 73 !" ~ = ()*ℎ 012 8 4 5" + 066 8 4 ℎ"#$ + 78 5" 9"      
  135.       !" !"#$ ℎ"#$ ×

    + × ()*ℎ × ()*ℎ + + + ℎ" ," -" !" ~ /" 0" ) ( -" = + 234 5 6 /" + 277 5 6 ℎ"#$ + 85 !" ~ = ()*ℎ 234 9 6 /" + 277 9 6 ℎ"#$ + 89
  136.       !" !"#$ ℎ"#$ ×

    + × ()*ℎ × ()*ℎ + + + ℎ" ," -" !" ~ /" 0" ) ( -" = + 234 5 6 /" + 277 5 6 ℎ"#$ + 85 !" ~ = ()*ℎ 234 9 6 /" + 277 9 6 ℎ"#$ + 89
  137.       !" !"#$ ℎ"#$ ×

    + × ()*ℎ × ()*ℎ + + + ℎ" ," -" !" ~ /" 0" ) ( -" = + 234 5 6 /" + 277 5 6 ℎ"#$ + 85 !" ~ = ()*ℎ 234 9 6 /" + 277 9 6 ℎ"#$ + 89
  138. ( ) !" !"#$ ℎ"#$ × + × ()*ℎ ×

    ()*ℎ + + + ℎ" ," -" !" ~ ," 0 ," ×!"#$ !"#$ !" 1 0 ," ×!"#$ !"#$ !" 1 /" 0" !" = ," ×!"#$ + -" ×!" ~
  139. ( ) !" !"#$ ℎ"#$ × + × ()*ℎ ×

    ()*ℎ + + + ℎ" ," -" !" ~ -" 0 -" ×!" ~ !" ~ /" !" 1 0 -" ×!" ~ !" ~ /" !" 1 /" 0" !" = ," ×!"#$ + -" ×!" ~
  140. ( ) !" !"#$ ℎ"#$ × + × ()*ℎ ×

    ()*ℎ + + + ℎ"   ," -" !" ~   ," -"    "!#"  /" 0" !" = ," ×!"#$ + -" ×!" ~
  141.     !" !" !"#$ ℎ"#$ × +

    × ()*ℎ × ()*ℎ + + + ℎ"   ," ," = + ./0 1 2 3" + .44 5 2 ℎ"#$ + 65 ℎ" = ," ×()*ℎ !"#$ 3" 7" % #   !   # $ "  7" = 89:(;)3(.=>? 1 2 ℎ" + 6=>? 1 )
  142.     !" !" !"#$ ℎ"#$ × +

    × ()*ℎ × ()*ℎ + + + ℎ" ," ," 0 1 ," ×()*ℎ !"#$ !" ℎ" 0 1 ," ×()*ℎ !"#$ !" ℎ" -" ." ," = + 012 3 4 -" + 055 6 4 ℎ"#$ + 76 ℎ" = ," ×()*ℎ !"#$ ." = 89:(;)-(0=>? 3 4 ℎ" + 7=>? 3 )
  143.     !" !" !"#$ ℎ"#$ × +

    × ()*ℎ × ()*ℎ + + + ℎ" ," -" ." -" ( ."/$ ) ( ." ) ( ," = + 123 4 5 -" + 166 7 5 ℎ"#$ + 87 ℎ" = ," ×()*ℎ !"#$ ." = 9:;(<)-(1>?@ 4 5 ℎ" + 8>?@ 4 )
  144.  c o h i B 4 O c a7

    O 5 3 - 12 . 6 l c n e c c