1 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models: Exploit Adversarial Examples with Graph-based Virtual Labels

발표일: 22.01.10 (월)

발표자: 황유진

Sejong RCV

KAIST RCV Lab.

2 of 14

This paper is to :

A graph-based SSL-AL framework

to unleash the SSL task models’ power and make an effective SSL-AL interaction.

​

​

Proposed famework:

gRaphbasEd VIrtual adVersarial Active Learning (REVIVAL)

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

Sejong RCV

KAIST RCV Lab.

3 of 14

What is SSL-AL:

​

Combine Semi-supervised Learning and Active Learning.

​

- Semi-supervised Learning:

Use labeled and unlabeled data for training a model

​

- Active Learning:

Select queries from unlabeled data to supplement labeled pool.

​

Both methods are used for mitigating the reliance on large labeled datasets.

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

4 of 14

  1. Their fully supervised task models are subjected to data-wasting problems.
  2. Most of these works are based on the VAE-GAN structure, which is learned through a mini-max game.

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Mini-max game 방식의 학습으로 야기되는 mismatch problem의 예시

Sejong RCV

KAIST RCV Lab.

5 of 14

  1. Their fully supervised task models are subjected to data-wasting problems.
  2. Most of these works are based on the VAE-GAN structure, which is learned through a mini-max game.

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Semi-supervised Task model

Graph-based structure

Sejong RCV

KAIST RCV Lab.

6 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

7 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

1. Label Propagator

​

Prior: Samples with their neighbors naturally form clusters in feature space, and samples belonging to the same cluster are more likely to share the same true labels.

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

8 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

2. Virtual Adversarial Generator

​

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

9 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

3. Boundary Limitator

​

queries

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

10 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

Datasets

​

  1. Image classification with CIFAR-10, CIFAR-100
  2. Segmentation with Cityscapes dataset.

​

Baseline

  1. ICAL (high inconsistency of predictions)
  2. Core-set (distribution-based)
  3. Entropy (uncertainty-based)
  4. Random

​

Supervised Active Learning

  1. SRAAL (State-Relabeling Adversarial Active Learning)
  2. VAAL
  3. LLAL (learning loss for active learning)
  4. CDAL

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

11 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

Image classification with CIFAR-10

the dotted line (p100) at the top represents the performance with the entire training set labeled.

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

12 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

Image classification with CIFAR-100

the dotted line (p100) at the top represents the performance with the entire training set labeled.

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Sejong RCV

KAIST RCV Lab.

13 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

Segmentation with Cityscapes

the dotted line (p100) at the top represents the performance with the entire training set labeled.

Sejong RCV

KAIST RCV Lab.

14 of 14

[ICCV 2021] Semi-supervised Active Learning for Semi-supervised Models

- This paper is to

- What is SSL-AL

- Problem of before work

- Method

- Experiments

- End

[CVPR 2020] State-Relabeling Adversarial Active Learning(SRAAL)

​

Sejong RCV

KAIST RCV Lab.