[DL輪読会]Taskonomy: Disentangling Task Transfer Learning
>100 Views
July 06, 18
スライド概要
2018/07/06
Deep Learning JP:
http://deeplearning.jp/seminar-2/
DL輪読会資料
関連スライド
各ページのテキスト
DEEP LEARNING JP [DL Papers] Taskonomy: (CVPR2018) Disentangling Task Transfer Learning Masashi Yokota, RESTAR inc. http://deeplearning.jp/ 1
• 2 CBCAL :BG B B 2 2 B : : BB • A 0 A : B : 1 A 1 :B :CB 8 , G:B . 1 C1 : : • 1G B C B : GL B : GL C CB : : :L • /0 : G / D: J • GGD G CBCAL CB
• • 3 = 2 36
2D Segm 2D Edge4 24 • E 4 4 2 2 4 -2>D
• 5 5
! 6
• 8 6 2
• 9 4 4 0 0
- • 1 - 0
• > 1 t: target task, s: source task Es: encoder, D: decoder, L: loss func, ft(I): grand truth • ( ) () )
• • 1 2 F - 2 ES 1 1 F 1 F 1 TE 2 2 > 1 > F 1
( ) ) :) • ) 1 1 3 ) A
) ( • ) • ( 4 4 r o ( ( !# 1 gu t1 e 1 )) ( s c 1 g !" a c 4 t L
) ( Source ( target 1 source s1 s2 s3 s4 s1 1 0.8 0.5 0.9 s2 0.2 1 0.7 0.5 s3 0.5 0.3 1 0.8 s4 0.1 0.5 0.2 1 s1 s2 s3 s4 s1 1 0.2 0.5 0.1 s2 0.8 1 0.3 0.5 s3 0.5 0.7 1 0.2 s4 0.9 0.5 0.8 1 5 W W s1 s2 s3 s4 s1 1/1 0.8/ 0.2 0.5/ 0.5 0.9/ 0.1 s2 0.2/ 0.8 1/1 0.7/ 0.3 0.5/ 0.5 s3 0.5/ 0.5 0.3/ 0.7 1/1 0.8/ 0.2 s4 0.1/ 0.9 1 0.2/ 0.8 1/1 W 5 s1 s2 s3 s4 s1 1 4 1 9 s2 0.25 1 7/3 1 s3 1 3/7 1 4 s4 1/9 1 0.25 1
) ( ( 6 6 1 Normalize s1 s2 s3 s4 s1 1 4 1 9 s1 0.47 t1 s2 0.25 1 7/3 1 s2 0.20 t2 s3 1 3/7 1 4 s3 0.25 t3 s4 1/9 1 0.25 1 s4 0.08 t4 PE 1 Target s1 s2 s3 s4 0.47 0.20 0.25 0.08
) ( ( target s1 s2 s3 s4 s1 s2 s3 s4 s1 s2 s3 s4 s1 1 s14 s21 s39 s4 s1 1 4 1 9 s1 1 4 1 9 s2 s1 0.25 1 1 47/3 1 1 9 s2 0.25 1 7/3 1 s2 0.25 1 7/3 1 s3 s21 0.25 3/7 1 1 7/34 1 s3 1 3/7 1 4 s3 1 3/7 1 4 s4 s31/9 1 1 3/7 0.25 1 1 4 s4 1/9 1 0.25 1 s4 1/9 1 0.25 1 s4 1/9 1 0.25 1 7 1 1 Normalize PE PE PE s1 0.47 PE s1 0.47 s1 0.47 s2 s1 0.20 0.47 s2 0.20 s2 0.20 s3 s2 0.25 0.20 s3 0.25 s3 0.25 s4 s3 0.08 0.25 s4 0.08 s4 0.08 s4 0.08 Target s1 s2 s3 s4 t1 0.47 0.20 0.25 0.08 t2 0.10 0.42 0.18 0.30 t3 0.21 0.11 0.33 0.35 t4 0.01 0.20 0.29 0.50
( ) • P 2 ) :) ) 8 P ) ) : 1 )0 ,), (
/ a e • • • I 1 9 1 2 / k 1 on /d c9
• 0 AHP target task
• V ( 45 7 45 35 D D • 2 ) 0 • ,2 2 5 – 28 – 12:842 8 – 5 ) 2 !" # 2 65 5 2 65 5 65 6 3 8 ( TP U 538683 5 2 65 6 3 8 ( G U UG UG :
• • • 2 2 2 DC 2 2 ( 2 :E 2 )
• % ( 3 G G : G • 2 : ) 3 G G G G
E : • 1 • 6 4 4 ( 1) 0 ( 2 2 1
: • 5 .5 D [ O Q . • ( 2 O : Q ). G 5 .5 G S
- - : • 2 6 6
) )) ( 7 7 7 2
& & N E & P & & I M • I 2M 8 N
• 2 .3 . • 6 0 • .3 6 6 3 6 3 3 0 6 0 6 2