Learning Loss for Active Learning Rymarczyk D., Zieliski B., Tabor - - PowerPoint PPT Presentation

▶
learning loss for active learning
SMART_READER_LITE
LIVE PREVIEW

Learning Loss for Active Learning Rymarczyk D., Zieliski B., Tabor - - PowerPoint PPT Presentation

Apr 05, 2024 366 likes •804 views

Learning Loss for Active Learning Rymarczyk D., Zieliski B., Tabor J., Sadowski M., Titov M. Agenda 1. Active Learning introduction 2. Base methods in AL for Deep Learning 3. Learning Loss for Active Learning 4. Our ideas for Active

slide-1
SLIDE 1

Learning Loss for Active Learning

Rymarczyk D., Zieliński B., Tabor J., Sadowski M., Titov M.

slide-2
SLIDE 2

Agenda

1. Active Learning introduction 2. Base methods in AL for Deep Learning 3. Learning Loss for Active Learning 4. Our ideas for Active Learning 5. Future plans 6. Bibliography

slide-3
SLIDE 3

Active Learning

Unlabeled dataset Labeled

slide-4
SLIDE 4

Active Learning

Unlabeled dataset Labeled Label

slide-5
SLIDE 5

Active Learning

Unlabeled dataset Labeled Predicted Label

slide-6
SLIDE 6

Active Learning

Unlabeled dataset Labeled Predicted Label True Label

slide-7
SLIDE 7

Active Learning

Unlabeled dataset Labeled Label

slide-8
SLIDE 8

Challenges in Active Learning

1. Criteria on which the sample will be chosen to the labelling process. 2. How many samples should be included in the labelling process? 3. Is the oracle infallible? 4. Multi oracle scenarios. 5. Online learning. 6. Can we use unlabeled data? and how? 7. How does the oracle reckon the AL system? 8. ...

slide-9
SLIDE 9

Challenges in Active Learning

1. Criteria on which the sample will be chosen to the labelling process. 2. How many samples should be included in the labelling process? 3. Is the oracle infallible? 4. Multi oracle scenarios. 5. Online learning. 6. Can we use unlabeled data? and how? 7. How does the oracle reckon the AL system? 8. ...

slide-10
SLIDE 10

Base methods in AL for Deep Learning

Random sampling to label

Unlabeled dataset Labeled

slide-11
SLIDE 11

Base methods in AL for Deep Learning

Core-Set approach - K-Greedy algorithm / KMeans++

Unlabeled dataset Labeled

slide-12
SLIDE 12

Base methods in AL for Deep Learning

Core-Set approach - K-Greedy algorithm / KMeans++

Unlabeled dataset Labeled

slide-13
SLIDE 13

Base methods in AL for Deep Learning

Uncertainty based approach - entropy

Unlabeled dataset

0.2 0.9 0.1 0.5 0.4 0.3 0.4

Labeled

0.1

slide-14
SLIDE 14

Base methods in AL for Deep Learning

Uncertainty based approach - entropy

Unlabeled dataset

0.2 0.9 0.1 0.5 0.4 0.3 0.4

Labeled

0.1

slide-15
SLIDE 15

Base methods in AL for Deep Learning

Experiment - learning episode:

Unlabeled dataset (CIFAR10) Labeled 1000 Predicted Label True Label To be Labeled 1000

slide-16
SLIDE 16

Base methods in AL for Deep Learning

Experiment - 10 x learning episodes:

Unlabeled dataset (CIFAR10) Labeled 10000 Predicted Label

slide-17
SLIDE 17

Base methods in AL for Deep Learning

Experiment - Results

slide-18
SLIDE 18

Learning Loss for Active Learning

Motivation: 1. None of the basic methods use information from inner layers of NN. 2. Best measure of NN error is value of loss function. 3. More advanced methods requires:

a. modifications of the architecture, b. training another neural network, c. training generative model, d. finding adversarial examples, e. bayesian deep learning, f. model ensembles.

slide-19
SLIDE 19

Learning Loss for Active Learning

Architecture modifications - learning loss module:

slide-20
SLIDE 20

Learning Loss for Active Learning

Architecture modifications - learning loss module:

slide-21
SLIDE 21

Learning Loss for Active Learning

Loss function for learning the loss

slide-22
SLIDE 22

Learning Loss for Active Learning

Results from the paper: Experiment details on CIFAR10:

  • network trained for 200 epochs, lr=0.1
  • after 160 epochs lr=0.01
  • at 120 epoch loss prediction module

does not influence the con weights

slide-23
SLIDE 23

Our ideas for Active Learning

  • 1. Remove the loss prediction module and use decoder or VAE.

Take to the labelling process samples with the highest reconstruction loss

slide-24
SLIDE 24

Our ideas for Active Learning

  • 1. Remove the loss prediction

module and use decoder or VAE.

slide-25
SLIDE 25

Our ideas for Active Learning

  • 2. We should try to make an adversarial example of the image and choose

those which requires the smallest modification to do so. DONE: https://arxiv.org/pdf/1802.09841.pdf

slide-26
SLIDE 26

Our ideas for Active Learning

  • 3. We should be like GANs. Train a discriminator to distinguish between

labeled and unlabeled dataset. DONE: https://arxiv.org/pdf/1907.06347.pdf

slide-27
SLIDE 27

Our ideas for Active Learning

  • 4. The neural network is learning the easy example first. Can be the history
  • f learning a differentiation between labeled and unlabeled datasets?
slide-28
SLIDE 28

Our ideas for Active Learning

  • 4. History of learning

Unlabeled dataset Labeled Label History record every 20 epochs r a w

slide-29
SLIDE 29

Our ideas for Active Learning

  • 4. History of learning

Unlabeled dataset Labeled History record RandomForest Labeled Unlabeled Only unlabeled 1000 100 sampled with highest possibility of being unlabeled

slide-30
SLIDE 30

Our ideas for Active Learning

  • 4. History of learning
slide-31
SLIDE 31

Our ideas for Active Learning

  • 5. Different moments of

History of learning

slide-32
SLIDE 32

Our ideas for Active Learning

  • 6. Maybe the NN is after

critical point - do not fine-tune

slide-33
SLIDE 33

Our ideas for Active Learning

  • 7. Take the history of inner layers
slide-34
SLIDE 34

Our ideas for Active Learning

  • 8. Why entropy is so good?

Can we be better?

slide-35
SLIDE 35

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-36
SLIDE 36

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-37
SLIDE 37

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-38
SLIDE 38

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-39
SLIDE 39

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-40
SLIDE 40

Our ideas for Active Learning

  • 9. Is history even worth something?
slide-41
SLIDE 41

Future plans

1. Investigate ways of finding the dataset outlier. 2. Do more research about history of learning.

slide-42
SLIDE 42

Future plans

1. IDEA: use augmentation for checking how the image prediction is sustained through different transformation. Take samples with highest m.

y’ y’’ y’’’

slide-43
SLIDE 43

Bibliography

1. Yoo, Donggeun, and In So Kweon. "Learning Loss for Active Learning." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019. https://arxiv.org/pdf/1905.03677.pdf 2. Ducoffe, Melanie, and Frederic Precioso. "Adversarial active learning for deep networks: a margin based approach." arXiv preprint arXiv:1802.09841 (2018). https://arxiv.org/pdf/1802.09841.pdf 3. Gissin, Daniel, and Shai Shalev-Shwartz. "Discriminative active learning." arXiv preprint arXiv:1907.06347 (2019). https://arxiv.org/pdf/1907.06347.pdf 4. Sener, Ozan, and Silvio Savarese. "Active learning for convolutional neural networks: A core-set approach." arXiv preprint arXiv:1708.00489 (2017). https://arxiv.org/pdf/1708.00489.pdf