[DL輪読会]Addressing Failure Prediction by Learning Model Confidence

DEEP LEARNING JP [DL Papers] Addressing Failure Prediction by Learning Model Confidence http://deeplearning.jp/ 1

• http://deeplearning.jp/

zt • • • • • . C A1 CM C LCH P 4 M3 8 , BLLI CN H A , ) BLLI - ACLBM HF N H C 0H v • 0B • 0H CS 0H L 9BHF .N / 2 5 LLBC M 0H ( LC T y • dfhbr • C H C A 5H n N H C 8H H : CN CLP ( k • ecg l • 90 ax m s 90 9 M 0 a i Uup H C CLP a i o 2

• https://arxiv.org/abs/1910.04851
• https://github.com/valeoai/ConfidNet

melb • x )) – – • f s ( – – – – ( ND a n p S – • t m idk - v )) t S c : .-. ! C u t t ODP ( (- = - P N M r - : .-. : -: - - = : .= t - .:P D ! P o PDSD - =:- y SD S ! t SD 3

• C – M P C M 4

• inh 11 D D g • ys • J8 8 DCprtg 8C ( • 0+2 gh – ys g – – +2 l v d 8 G a g g V 11 C8 8 8C8C u N g ad 8 6 G D 7 ( [ v Nm d D ( d 11 d . 8C TMk o + 8GG 2 D989 M g g )6 dbP [M G D 7 g V o d + +D 9 L sN g d, c]e G S e d ) 5

• • • • • !: #∗ #% & ' 6

• – T – T – C • T M • P : • :P : 7

2 • - P – – – ) ( 2 2 - - 2 N L 2 2 2 2 C T E = E 2 2 - - E E 8

ea • VMSRT – glm) 1 0 – RP UM • V h – lm) – RP • 9 ,0. 5 % ,0. 5 %% ) ,:B9 c UM FI H F ) c H FI H : :C ,A: F OiN H OiN d – – 5 -FF F 5 5) H :F : IC I FH . 5 :H ( F C 5 5 5 , :AA IF C I C FF F : H E H A: 9

• 10

• 11

• • • N N C N 12

J • 7 DDD :834 70 4 4 03 4 0 80: 0 029 44 .40 8 6-/ 3:0 8 • 7 DDD :834 70 4 4 44 .40 8 6-/ 3:82: • 7 0 8 6 35 35 • 7 0 8 6 35 35 • 7 0 8 6 35 35 • 7 0 8 6 35 35 • 7 0 8 6 35 :4 8 43 5 0 4D 9 5 34 42 8 6 538 81 8 0 :4 0 3 13

• https://www.slideshare.net/DeepLearningJP2016/dla-simple-unified-framework-for-detecting-outofdistribution-samples-and-adversarial-attacks-118803897
• https://www.slideshare.net/DeepLearningJP2016/dliclr2020
• https://arxiv.org/pdf/1506.02142.pdf
• https://arxiv.org/pdf/1706.04599.pdf
• https://arxiv.org/pdf/1612.01474.pdf
• https://arxiv.org/pdf/1805.11783.pdf
• https://arxiv.org/pdf/1706.02690

14

/ • /1 /5: : 65 B 5 /1 /: 8 2 :/ : 8/ A 2:: 8 681 .0 .0 15

• https://arxiv.org/pdf/1511.02680.pdf

P • T – – • : C : : ) ( B E M 16

N • C F ) : : ) ( ) 17