[DL輪読会]Addressing Failure Prediction by Learning Model Confidence
1 like435 views
1. This document discusses several papers on deep learning failure prediction and model confidence. It covers techniques for predicting when deep learning models may fail and estimating confidence in their predictions. 2. The document also discusses approaches for improving model robustness, such as data augmentation and regularizing models to be less sensitive to perturbations in inputs. 3. Additionally, it briefly outlines methods for visualizing what deep learning models have learned, including saliency maps and class activation mapping.
![1
DEEP LEARNING JP
[DL Papers]
http://deeplearning.jp/
Addressing Failure Prediction by Learning Model Confidence](https://image.slidesharecdn.com/dl20191101nishimura-191101000858/85/DL-Addressing-Failure-Prediction-by-Learning-Model-Confidence-1-320.jpg)
![• inh 11 D D g V M 11 g ad 8 6 G D 7
• ys gh l v d 8 G C8 8 8C8C (
• J8 8 DCprtg a u N [ v Nm d D (
8C (
• 0+2 g d
– ys 11 g dbP [M sN c]e S e d
g d . 8C ) 6 G D 7
– g TMk o g V o g d , G )
– +2 + 8GG 2 D989 g d + +D 9 L
5](https://image.slidesharecdn.com/dl20191101nishimura-191101000858/85/DL-Addressing-Failure-Prediction-by-Learning-Model-Confidence-5-320.jpg)
[DL輪読会]Addressing Failure Prediction by Learning Model Confidence
- 1. 1 DEEP LEARNING JP [DL Papers] http://deeplearning.jp/ Addressing Failure Prediction by Learning Model Confidence
- 2. zt • . C A 1 C M C LCH P 4 C A 5H 0H • M3 8 , • BLLI - CN H A , ) • BLLI - ACLBM HF N H C 0H L • v • 0B 0H CS C H 9BHF .N / 2 5 LLBC M 0H ( L C T • y • dfhbr n N H C 8H H : CN CLP ( • k • ecg l s 90 9 M 0 H C CLP a i • 90 ax m a i Uup o 2
- 3. melb m • x )) t ND – s a n v )) t S c C – p u • f S idk t ODP ( (- = - : .-. ! P N – - t M • ( r – - : .-. - .:- PDSD - =:- ! – - : -: - - - P D – = : .= ! P y SD – ( t o S t SD 3
- 4. • C M – P C M 4
- 5. • inh 11 D D g V M 11 g ad 8 6 G D 7 • ys gh l v d 8 G C8 8 8C8C ( • J8 8 DCprtg a u N [ v Nm d D ( 8C ( • 0+2 g d – ys 11 g dbP [M sN c]e S e d g d . 8C ) 6 G D 7 – g TMk o g V o g d , G ) – +2 + 8GG 2 D989 g d + +D 9 L 5
- 6. • !: • #∗ • %# • & • ' 6
- 7. • – T • – T – C T M • P : • :P : 7
- 8. 2 - • - P ) ( T E 2 - E – 2 2 - E – 2 2 - 2 - C = 2 - E – 2 - N L 2 8
- 9. ea • VMSRT – glm) 1 0 9 ,0. 5 % ,0. 5 %% – RP UM ) ,:B9 • V h – lm) c FI H F H FI H ,A: F OiN – RP UM ) c : :C H OiN • d – 5 -FF F 5 I . 5 :H ( 5 5 , – 5) H :F : IC F H F C 5 :AA IF C I C FF F : H E H A: 9
- 10. • 10
- 11. • 11
- 12. • N • N C • N 12
- 13. J • 7 DDD :834 70 4 4 44 .40 8 6-/ 3:0 8 :4 8 43 5 0 4D 9 5 34 42 8 6 538 81 8 0 :4 0 3 03 4 0 80: 0 029 • 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 13
- 14. 14
- 15. / 5 /1 /: 8 2 :/ : 8/ • /1 /5: : 65 B A 15 2:: 8 681 .0 .0
- 16. P : C : • T : – ) ( – B • E M 16
- 17. N ) ) ( • : - C F : ) 17