Presentation is loading. Please wait.

Presentation is loading. Please wait.

Laso: Label-Set Operations Networks for Multi-label Few-shot Learning

Similar presentations


Presentation on theme: "Laso: Label-Set Operations Networks for Multi-label Few-shot Learning"— Presentation transcript:

1 Laso: Label-Set Operations Networks for Multi-label Few-shot Learning
CVPR 2019 oral Laso: Label-Set Operations Networks for Multi-label Few-shot Learning Amit Alfassy, Leonid Karlinsky, Amit Aides IBM Research AI Haifa, Israel Fewshot: Only a handful of examples are available for the task

2 A story: The Tank Recognition Project
In 1980, the America government started a project for tank recognition.

3 A story: The Tank Recognition Project
In 1980, the America government started a project for tank recognition. “0” “1”

4 A story: The Tank Recognition Project
In 1980, the America government started a project for tank recognition. “0” “1” The accuracy on the test dataset was high, but when it turned into real world scenarios, it did not work at all !

5 A story: The Tank Recognition Project
In 1980, the America government started a project for tank recognition. Why? “0” “1” The accuracy on the test dataset was high, but when it turned into real world scenarios, it did not work at all !

6 A story: The Tank Recognition Project
In 1980, the America government started a project for tank recognition. “0” “1” Sunny Rainy

7 A story: The Tank Recognition Project
Machine Learning Researcher: It’s overfitting! Apply data augmentation、collect more images from different weather condition

8 A story: The Tank Recognition Project
Researcher of Interpretability of NN: The recognition process is a black-box! Visualize the locations that fire on the recognition task

9 A story: The Tank Recognition Project
After I read this paper: The tank is coupled with the weather! De-couple them from each other!

10 A story: The Tank Recognition Project
After I read this paper: The tank is coupled with the weather! De-couple them from each other! In multi-label few-shot learning, objects may easily coupled from each other, resulting in undesirable results.

11 Methodology

12 Methodology 𝒁 𝒊𝒏𝒕 = 𝑴 𝒊𝒏𝒕 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝒁 𝒖𝒏𝒊 = 𝑴 𝒖𝒏𝒊 ( 𝑭 𝒙 , 𝑭 𝒚 )
𝒁 𝒔𝒖𝒃 = 𝑴 𝒔𝒖𝒃 ( 𝑭 𝒙 , 𝑭 𝒚 )

13 Methodology 𝐶 𝑙𝑜𝑠𝑠 𝒁 𝒊𝒏𝒕 = 𝑴 𝒊𝒏𝒕 ( 𝑭 𝒙 , 𝑭 𝒚 )
𝒁 𝒖𝒏𝒊 = 𝑴 𝒖𝒏𝒊 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝒁 𝒔𝒖𝒃 = 𝑴 𝒔𝒖𝒃 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝐶 𝑙𝑜𝑠𝑠

14 Methodology 𝐿𝑎𝑠𝑜 𝑙𝑜𝑠𝑠 𝒁 𝒊𝒏𝒕 = 𝑴 𝒊𝒏𝒕 ( 𝑭 𝒙 , 𝑭 𝒚 )
𝒁 𝒖𝒏𝒊 = 𝑴 𝒖𝒏𝒊 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝒁 𝒔𝒖𝒃 = 𝑴 𝒔𝒖𝒃 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝐿𝑎𝑠𝑜 𝑙𝑜𝑠𝑠

15 Methodology 𝑅 𝑙𝑜𝑠𝑠 𝑠𝑦𝑚 𝒁 𝒊𝒏𝒕 = 𝑴 𝒊𝒏𝒕 ( 𝑭 𝒙 , 𝑭 𝒚 )
𝒁 𝒖𝒏𝒊 = 𝑴 𝒖𝒏𝒊 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝒁 𝒔𝒖𝒃 = 𝑴 𝒔𝒖𝒃 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝑅 𝑙𝑜𝑠𝑠 𝑠𝑦𝑚

16 Methodology 𝑅 𝑙𝑜𝑠𝑠 𝑚𝑐 𝒁 𝒊𝒏𝒕 = 𝑴 𝒊𝒏𝒕 ( 𝑭 𝒙 , 𝑭 𝒚 )
𝒁 𝒖𝒏𝒊 = 𝑴 𝒖𝒏𝒊 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝒁 𝒔𝒖𝒃 = 𝑴 𝒔𝒖𝒃 ( 𝑭 𝒙 , 𝑭 𝒚 ) 𝑅 𝑙𝑜𝑠𝑠 𝑚𝑐

17 Experiments Coco: object-based multi-label dataset
Celeb-A: attribute-based multi-label dataset(40 facial attributes)

18 Experiments Coco: object-based multi-label dataset
Celeb-A: attribute-based multi-label dataset (40 facial attributes) The Coco dataset is split into Coco-A(64 classes), Coco-B(16 classes) 𝐹 𝐴 , 𝑙𝑎𝑠𝑜 𝐴 , 𝐶 𝐴 𝐹 𝐴 , 𝑙𝑎𝑠𝑜 𝐴 , 𝐶 𝐵 Numbers are in mAP Numbers are in mIoU

19 Experiments Coco: object-based multi-label dataset
Celeb-A: attribute-based multi-label dataset (40 facial attributes)

20 Experiments: Ablations
Coco: object-based multi-label dataset Celeb-A: attribute-based multi-label dataset (40 facial attributes) Tab. Analytic approximations to set operations

21 Experiments: Ablations
Coco: object-based multi-label dataset Celeb-A: attribute-based multi-label dataset (40 facial attributes) Tab. Analytic approximations to set operations

22 Experiments: Ablations
Coco: object-based multi-label dataset Celeb-A: attribute-based multi-label dataset (40 facial attributes) Tab. Serve as data augmentation strategy numbers are in mAP Conduct on Coco-B

23 Conclusion + First paper to address the multi-label classification of few-shot learning. + The Motivation —— which I would call it “de-couple” is novel. + Abundant experiments, well written and easy to follow. - The implementation of de-couple is disappointing.


Download ppt "Laso: Label-Set Operations Networks for Multi-label Few-shot Learning"

Similar presentations


Ads by Google