from: CVPR-15: Learning from Massive Noisy Labeled Data for Image Classification
- 14 classes: T-shirt, Shirt, Knitwear, Chiffon, Sweater, Hoodie, Windbreaker, Jacket, Down Coat, Suit, Shawl, Dress, Vest, and Underwear
- noisy labeled training dataset (
$D_\eta$ ):$10^6$ - clean train data(
$D_c$ ): 47,570 - clean validation set: 14,313
- clean test set: 10,526
It's not column-diagonally dominant, thus small-loss trick may not work. But if examples in noisy class 3 and noisy class 5 have been swapped, it may become column-diagonally dominant, in which case small-loss trick may work.
69.9(only use noisy training data) -> 79.9(fine-tuning)
10. CVPR-19: MetaCleaner: Learning to Hallucinate Clean Representations for Noisy-Labeled Visual Recognition
about 71%
10. NIPS-19: L_DMI: A Novel Information-theoretic Loss Function for Training Deep Nets Robust to Label Noise
from: Arxiv17: Webvision database: Visual learning and understanding from web data
- 1,000 classes: concepts in ImageNet ILSVRC12
- noise rate: 20-40%
3. ICML-18: MentorNet: Learning Data-Driven Curriculum for Very Deep Neural Networks on Corrupted Labels
use all data or only use the first 50 classes of Google image subset
use all 1000 classes
only use the first 50 classes of Google image subset
only use the randomly selected 100 classes
3. Food101-N
from: CVPR-18: CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise
- 101 food classes
- 310k image, 55k verification,
- noise rate: 20%
not use Food101N created by cleanNet paper, but use Food101 and inject 20% noise
3. CVPR-19: MetaCleaner: Learning to Hallucinate Clean Representations for Noisy-Labeled Visual Recognition
4. ANIMAL10N
from: ICML-19: SELFIE: Refurbishing Unclean Samples for Robust Deep Learning