from synthetic data to the holy grail of ai€¦ · sensor time series analysis failure prediction...

73
From synthetic data to the holy grail of AI GENERATIVE MODELS

Upload: others

Post on 23-Sep-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

From synthetic data to the holy grail of AIGENERATIVE MODELS

Page 2: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

2

ABOUT ME

- Solution Architect @ NVIDIA - Supporting delivery of AI / Deep Learning solutions

- 10 years experience working on HPC (high performance computing) systems, signal processing, and Machine Learning

- My past experience:

- Northrop Grumman – Systems Engineer

- Exxon Mobil – HPC Programmer

Aleksandr Volkov – [email protected]

Page 3: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

3

WHY DO NEURAL NETWORKS WORK?

Page 4: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

4

NEURAL NETWORKS ARE NOT NEWAnd are disapintingly simple as an algorithm

Page 5: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

5

SUPERVISED LEARNING

y= f(x)

Approximating complex functions

Page 6: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

6

SMALL NEURAL NETWORKSUnderperform

Data

Perfo

rman

ce

20101990 2000

Page 7: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

7

WHY?

Page 8: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

8

DATA

Data

Perfo

rman

ce

20101990 2000 2017 2020

Page 9: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

9

COMPUTEHistorically we never had large datasets or compute

1980 1990 2000 2010 2020

102

103

104

105

106

107

Single-threaded perf

1.5X per year

1.1X per yearTransistors(thousands)

GPU-Computing perf1.5X per year 1000X

By 2025

Page 10: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

10

PERSPECTIVE

Page 11: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

11

CONTEXT1.759 petaFLOPs in November 2009

Page 12: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

12

CONTEXT2 petaFLOPs - today

Page 13: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

13

CONTEXT TODAY200 petaFLOPs in November 2018

Achitecture:

9,216 POWER9 22-core CPUs

27,648 Nvidia Tesla V100 GPUs

Approaching 3.3 exaopsusing mixed precision

Page 14: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

14

2016 – Baidu Deep Speech 2Superhuman Voice Recognition

2015 – Microsoft ResNetSuperhuman Image Recognition

2017 – Google Neural Machine TranslationNear Human Language Translation

100 ExaFLOPS8700 Million Parameters

20 ExaFLOPS300 Million Parameters

7 ExaFLOPS60 Million Parameters

To Tackle Increasingly Complex ChallengesNEURAL NETWORK COMPLEXITY IS EXPLODING

Page 15: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

15

100 EXAFLOPS=

2 YEARS ON A DUAL CPU SERVER

Page 16: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

16

NEURAL NETWORKS ARE NOT NEWRequire abundance of data and compute

Data

Perfo

rman

ce

20101990 20172000 2020

Page 17: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

17

EXPLODING DATASETS

Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., ... & Zhou, Y. (2017). Deep Learning Scaling is Predictable, Empirically. arXiv preprint arXiv:1712.00409.

Logarithmic relationship between the dataset size and accuracy• Translation

• Language Models

• Character Language Models

• Image Classification

• Attention Speech Models

Page 18: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

18

EXPLODING DATASETS

Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., ... & Zhou, Y. (2017). Deep Learning Scaling is Predictable, Empirically. arXiv preprint arXiv:1712.00409.

Logarithmic relationship between the dataset size and accuracy

Page 19: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

19

MAKING COMPLEX PROBLEMS EASY

Page 20: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

20

TRANSFORMING IMPOSSIBLE INTO EXPENSIVE

Page 21: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

21

Page 22: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

22

SUPERVISED LEARNING

y= f(x)

Approximating complex functions

Page 23: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

23

HUGE OPPORTUNITY

Page 24: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

24

HUGE CHALLENGE

Page 25: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

25

EXPLODING DATASETS

Hestness, J., Narang, S., Ardalani, N., Diamos, G., Jun, H., Kianinejad, H., ... & Zhou, Y. (2017). Deep Learning Scaling is Predictable, Empirically. arXiv preprint arXiv:1712.00409.

Logarithmic relationship between the dataset size and accuracy

Page 26: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

26

A BETTER WAY

Page 27: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

27

GENERATIVE MODELS

y= f(x)

Reformulating the problem

Training setGenerate new data with the same statistics

Page 28: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

28

WE NO LONGER TRAIN OUR MODEL TO SOLVE A NARROW TASK

Page 29: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

29

WE TRAIN IT TO LEARN A PHENOMENON (ITS PROBABILITY DISTRIBUTION)

Page 30: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

30

TEACHING MACHINES HOW TO COMPOSE HUMAN LANGUAGE

Page 31: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

31

UNDERSTANDING LANGUAGEGenerating text

Transformer LM – 100M Parameters

Page 32: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

32

UNDERSTANDING LANGUAGEGenerating text

Transformer LM – 5B Parameters

Page 33: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

33

“Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT-2 along with sampling code. We are not releasing the dataset, training code, or GPT-2 model weights.”

“Better Language Models and Their Implications”, OpenAI

Page 34: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

34

TEACHING MACHINES HOW TO COMPOSE MUSIC

Page 35: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

35

UNDERSTANDING BEAUTYComposing music

Page 36: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

36

TEACHING MACHINES HOW TO GENERATE CODE

Page 37: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

37

UNDERSTANDING CODEGenerating computer code

Page 38: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

38

TEACHING MACHINES HOW TO GENERATE OTHER FORMS OF HUMAN LANGUAGE

Page 39: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

39

UNDERSTANDING DESIGN2D drawings to 3D sketches

Yao, Jiaxian, et al. "Interactive Design and Stability Analysis of Decorative Joinery for Furniture." ACM Transactions on Graphics (TOG) 36.2 (2017): 20.

Page 40: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

40

UNDERSTANDING DESIGNModels from specification

Autodesk research

Page 41: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

41

TEACHING MACHINES HOW TO SPEAK

Page 42: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

42

UNDERSTANDING SPEECHThe presence of speech generation

http://deepzen.io/

Page 43: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

43

TEACHINGS MACHINES HOW TO SING

Page 44: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

44

UNDERSTANDING BEAUTYPerforming music

Page 45: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

45

UNDERSTANDING BEAUTYPerforming music

Page 46: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

46

TEACHINGS MACHINES ABOUT HUMANS

Page 47: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

47

UNDERSTANDING IMAGENeural Network is computing most likely value of pixels

Sønderby, Casper Kaae, et al. "Amortised map inference for image super-resolution." arXiv preprint arXiv:1610.04490 (2016).

Page 48: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

48

UNDERSTANDING THE SHAPE OF A FACEFrom a single image

Saito, Shunsuke, et al. "Photorealistic Facial Texture Inference Using Deep Neural Networks." arXiv preprint arXiv:1612.00523 (2016).

Page 49: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

49

UNDERSTANDING A CONCEPT OF A FACEGeneration

Page 50: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

50

ANIMATING A FACEAnd more

Karras, Tero, et al. "Audio-driven facial animation by joint end-to-end learning of pose and emotion." ACM Transactions on Graphics (TOG) 36.4 (2017): 94.

Page 51: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

51

ANIMATING A FACETeaching the network to animate facial expressions

Page 52: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

52

TEACHING MACHINES ABOUT MOTION

Page 53: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

53

CHARACTER ANIMATIONTeaching the network to animate movement

Page 54: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

54

CHARACTER ANIMATIONTeaching the network to animate movement

Page 55: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

55

TEACHINGS MACHINES ABOUT THE VISUAL WORLD

Page 56: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

56

UNDERSTANDING IMAGETeaching the network the physics of the world

Page 57: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

57

UNDERSTANDING IMAGE QUALITY

http://people.ee.ethz.ch/~ihnatova/

Page 58: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

58

UNDERSTANDING THE VISUAL PROPERTIES OF THE WORLD

Page 59: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

59

UNDERSTANDING THE VISUAL PROPERTIES OF THE WORLD

Page 60: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

60

UNDERSTANDING IMAGE GENERAL PROPERTIES

How should the image look like from a different angle

Page 61: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

61

TEACHING MACHINES ABOUT STYLE AND BEAUTY

Page 62: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

62

SEMANTIC STYLE TRANSFERUnderstanding common semantic part of the image

Page 63: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

63

GENERATING IMAGES FROM DESCRIPTION

Reed, Scott, et al. "Parallel Multiscale Autoregressive Density Estimation." arXiv preprint arXiv:1703.03664 (2017).

Page 64: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

64

NOT JUST THEORETICAL RESEARCHNvidia GameWorks

Page 65: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

65

TEACHINGS MACHINES ABOUT THE PHYSICAL WORLD

Page 66: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

66

AI – THE NEW INSTRUMENT FOR SCIENCENeed for general purpose compute

Page 67: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

67

NEURAL NETWORKS FOR SCIENCE

AI-POWERED WEATHER PREDICTION

PLASMA FUSION APPLICATION EARTHQUAKE SIMULATION

7.815.7

125

0

20

40

60

80

100

120

140

V100 TFLOPS

FP64+ MULTI-PRECISION

FP16 Solver3.5x times faster

FP16/FP321.15x ExaOPS

FP16-FP21-FP32-FP6425x times faster

Page 68: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

68

AI – THE NEW INSTRUMENT FOR SCIENCEFluid dynamics

Page 69: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

69

TEACHINGS MACHINES ABOUT ECONOMY

Page 70: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

70

UNDERSTANDING ECONOMICSGenerating stock movement

Page 71: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

71

• Usually postings or documents that exhibit an unusual or rare attribute values, such as:

• Seldom used user accounts,• Reverse postings, corrections

„Global“ Accounting Anomalies

• Usually postings or documents that exhibit an unusual or rare attribute combination, for example:

• Unusual posting activities• Deviating user behavior

„Local“ Accounting Anomalies#

Feat

ure

2 (e

.g. L

ine-

Item

s)

# Feature 1 (e.g. Amount)

# Fe

atur

e 2

(e.g

. Lin

e-Ite

ms)

# Feature 1 (e.g. Posting Amount)

Understanding BusinessDifferentiating mistakes from fraud

Page 72: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

72

TEACHINGS MACHINES ABOUT MACHINES

Page 73: From synthetic data to the holy grail of AI€¦ · Sensor Time Series Analysis Failure Prediction Wide range of applications. Title: Alex Volkov - Generative models From synthetic

73

PREDICTIVE MAINTENANCEFIELD INSPECTIONFACTORY INSPECTION

AI FOR INDUSTRIAL APPLICATIONS

Quality InspectionFault Detection & Classification

Inventory Inspection

Condition Based MaintenanceRemaining Useful Life

Sensor Time Series AnalysisFailure Prediction

Wide range of applications