Contenu connexe Similaire à [DLHacks 実装]Network Dissection: Quantifying Interpretability of Deep Visual Representations (20) Plus de Deep Learning JP (20) [DLHacks 実装]Network Dissection: Quantifying Interpretability of Deep Visual Representations1. 1
DEEP LEARNING JP
[DL Papers]
http://deeplearning.jp/
“Network Dissection: Quantifying Interpretability of Deep
Visual Representations (CVPR2017)”
Yosuke Ueno
5. Goal: From Visualization to Interpretation
5
Interpretation: lamp
Interpretation: car
Score: 0.15
Score: 0.02
TopActivated Images
Unit 1
Unit 4
TopActivated Images
著者のスライドより引用
9. IoU (Intersection over Union)
9
• 例えば、馬については
(の合
計)
• x:データセットの画像
• k:チャネル
• 𝑀 𝑘:チャネルkの特徴マップの活発な部分
• 𝐿 𝑐:クラスcのground truth
10. Approach: Test units for semantic segmentation
10
Lamp Intersection over Union (IoU)= 0.12
Unit 1 Top activated images
著者のスライドより引用
13. 結果 (AlexNet trained on places365)
13
Histogram of object detectors: Detector:81/256, Unique Detector:40 (Units with IoU>0.04)
Living room
Kitchen
Coast
…
365 categories
conv5, 256 units 著者のスライドより引用