Generative Modeling by Estimating Gradients of the Data Distribution - - PowerPoint PPT Presentation

generative modeling by estimating gradients of the data
SMART_READER_LITE
LIVE PREVIEW

Generative Modeling by Estimating Gradients of the Data Distribution - - PowerPoint PPT Presentation

Generative Modeling by Estimating Gradients of the Data Distribution Yang Song, Stefano Ermon David Zimmerer Medical Image Analysis ( #MIA-san-mia ) DKFZ Author Division Generative modeling 02.11.16 | Page2 2 | Author Division


slide-1
SLIDE 1

Generative Modeling by Estimating Gradients of the Data Distribution

David Zimmerer Medical Image Analysis ( #MIA-san-mia ) DKFZ Yang Song, Stefano Ermon

slide-2
SLIDE 2

Page2 02.11.16 | Author Division

Generative modeling

2 |

slide-3
SLIDE 3

Page3 02.11.16 | Author Division

Generative modeling

3 |

slide-4
SLIDE 4

Page4 02.11.16 | Author Division

Generative modeling

4 |

slide-5
SLIDE 5

Page5 02.11.16 | Author Division

Generative modeling

5 |

slide-6
SLIDE 6

Page6 02.11.16 | Author Division

Generative modeling

6 |

slide-7
SLIDE 7

Page7 02.11.16 | Author Division

Generative modeling

7 |

slide-8
SLIDE 8

Page8 02.11.16 | Author Division

Generative modeling

8 |

slide-9
SLIDE 9

Page9 02.11.16 | Author Division

“New” Idea: Generative Modeling by Estimating Gradients of the Data Distribution

9 |

slide-10
SLIDE 10

Page10 02.11.16 | Author Division

“New” Idea: Generative Modeling by Estimating Gradients of the Data Distribution

10 |

slide-11
SLIDE 11

Page11 02.11.16 | Author Division

“New” Idea: Generative Modeling by Estimating Gradients of the Data Distribution

11 |

slide-12
SLIDE 12

Page12 02.11.16 | Author Division

“New” Idea: Generative Modeling by Estimating Gradients of the Data Distribution

12 |

slide-13
SLIDE 13

Page13 02.11.16 | Author Division

Score matching

13 |

[1]

  • A. Hyvärinen. Estimation of non-normalized statistical models by score matching. Journal of

Machine Learning Research, 6(Apr):695–709, 2005.

slide-14
SLIDE 14

Page14 02.11.16 | Author Division

Score matching

14 |

[1]

  • A. Hyvärinen. Estimation of non-normalized statistical models by score matching. Journal of

Machine Learning Research, 6(Apr):695–709, 2005.

slide-15
SLIDE 15

Page15 02.11.16 | Author Division

Score matching

15 |

[1]

  • A. Hyvärinen. Estimation of non-normalized statistical models by score matching. Journal of

Machine Learning Research, 6(Apr):695–709, 2005.

slide-16
SLIDE 16

Page16 02.11.16 | Author Division

Score matching

16 |

[1]

  • A. Hyvärinen. Estimation of non-normalized statistical models by score matching. Journal of

Machine Learning Research, 6(Apr):695–709, 2005.

slide-17
SLIDE 17

Page17 02.11.16 | Author Division

Score matching

17 |

→ →

[1]

  • A. Hyvärinen. Estimation of non-normalized statistical models by score matching. Journal of

Machine Learning Research, 6(Apr):695–709, 2005.

slide-18
SLIDE 18

Page18 02.11.16 | Author Division

Trivial Implementation (on MNIST)

18 |

slide-19
SLIDE 19

Page19 02.11.16 | Author Division

Trivial Implementation (on MNIST)

19 |

slide-20
SLIDE 20

Page20 02.11.16 | Author Division

Trivial Implementation (on MNIST)

20 |

slide-21
SLIDE 21

Page21 02.11.16 | Author Division

Spoiler: Improved Results

21 |

slide-22
SLIDE 22

Page22 02.11.16 | Author Division

Spoiler: Improved Results

22 |

slide-23
SLIDE 23

Page23 02.11.16 | Author Division

Spoiler: Improved Results

23 |

slide-24
SLIDE 24

Page24 02.11.16 | Author Division

Spoiler: Improved Results

24 |

slide-25
SLIDE 25

Page25 02.11.16 | Author Division

What are the problems ?

25 |

slide-26
SLIDE 26

Page26 02.11.16 | Author Division

What are the problems ?

26 |

slide-27
SLIDE 27

Page27 02.11.16 | Author Division

What are the problems ?

27 |

slide-28
SLIDE 28

Page28 02.11.16 | Author Division

What are the problems ?

28 |

slide-29
SLIDE 29

Page29 02.11.16 | Author Division

What are the problems ?

29 |

slide-30
SLIDE 30

Page30 02.11.16 | Author Division

What are the problems ?

30 |

slide-31
SLIDE 31

Page31 02.11.16 | Author Division

What are the problems ?

31 |

→ →

slide-32
SLIDE 32

Page32 02.11.16 | Author Division

How to sample:

32 |

slide-33
SLIDE 33

Page33 02.11.16 | Author Division

How to sample:

33 |

slide-34
SLIDE 34

Page34 02.11.16 | Author Division

How to sample:

34 |

slide-35
SLIDE 35

Page35 02.11.16 | Author Division

How to sample:

35 |

slide-36
SLIDE 36

Page36 02.11.16 | Author Division

How to sample:

36 |

slide-37
SLIDE 37

Page37 02.11.16 | Author Division

How to sample:

37 |

slide-38
SLIDE 38

Page38 02.11.16 | Author Division

Approach

38 |

slide-39
SLIDE 39

Page39 02.11.16 | Author Division

Approach

39 |

slide-40
SLIDE 40

Page40 02.11.16 | Author Division

Results

40 |

slide-41
SLIDE 41

Page41 02.11.16 | Author Division

Results: Qualitative

41 |

slide-42
SLIDE 42

Page42 02.11.16 | Author Division

Results: Qualitative

42 |

slide-43
SLIDE 43

Page43 02.11.16 | Author Division

Results: Quantitative

43 |

slide-44
SLIDE 44

Page44 02.11.16 | Author Division

Results: Reproducible

44 |

[2]

  • A. Matosevic. Reproducibility Challenge – Generative Modeling by Estimating Gradients of the Data Distribution,

NeurIPS 2019 Reproducibility Challenge Blind Report, https://openreview.net/forum?id=SkxCSTqG6H

slide-45
SLIDE 45

Page45 02.11.16 | Author Division

Results: Reproducible

45 |

[2]

  • A. Matosevic. Reproducibility Challenge – Generative Modeling by Estimating Gradients of the Data Distribution,

NeurIPS 2019 Reproducibility Challenge Blind Report, https://openreview.net/forum?id=SkxCSTqG6H

slide-46
SLIDE 46

Page46 02.11.16 | Author Division

Results: Reproducible

46 |

[2]

  • A. Matosevic. Reproducibility Challenge – Generative Modeling by Estimating Gradients of the Data Distribution,

NeurIPS 2019 Reproducibility Challenge Blind Report, https://openreview.net/forum?id=SkxCSTqG6H

slide-47
SLIDE 47

Page47 02.11.16 | Author Division

Results: Reproducible

47 |

[2]

  • A. Matosevic. Reproducibility Challenge – Generative Modeling by Estimating Gradients of the Data Distribution,

NeurIPS 2019 Reproducibility Challenge Blind Report, https://openreview.net/forum?id=SkxCSTqG6H

slide-48
SLIDE 48

Page48 02.11.16 | Author Division

Results: Reproducible

48 |

[2]

  • A. Matosevic. Reproducibility Challenge – Generative Modeling by Estimating Gradients of the Data Distribution,

NeurIPS 2019 Reproducibility Challenge Blind Report, https://openreview.net/forum?id=SkxCSTqG6H

slide-49
SLIDE 49

Page49 02.11.16 | Author Division

The End

49 |