SLIDE 1
THE MAGIC OF UNSUPERVISED LEARNING Agustinus Nalwan
Head of AI
Carsales.com.au
THE MAGIC OF UNSUPERVISED LEARNING Agustinus Nalwan Head of AI - - PowerPoint PPT Presentation
THE MAGIC OF UNSUPERVISED LEARNING Agustinus Nalwan Head of AI - - PowerPoint PPT Presentation
THE MAGIC OF UNSUPERVISED LEARNING Agustinus Nalwan Head of AI Carsales.com.au A LITTLE BIT ABOUT MYSELF AI 4 Years SAVING OUR FOREST BUSHFIRE DETECTION BACK TO THE MAIN TOPIC THE MAGIC OF UNSUPERVISED LEARNING SUPERVISED LEARNING
SLIDE 2
SLIDE 3
SAVING OUR FOREST
BUSHFIRE DETECTION
SLIDE 4
SLIDE 5
BACK TO THE MAIN TOPIC
SLIDE 6
THE MAGIC OF UNSUPERVISED LEARNING
SLIDE 7
SUPERVISED LEARNING
SLIDE 8
SUPERVISED LEARNING IMAGE RECOGNITION
SLIDE 9
SLIDE 10
SLIDE 11
CAR RECOGNITION
SLIDE 12
SLIDE 13
Ford Kuga Titanium
SLIDE 14
I am awesome
SUPERVISED LEARNING
SLIDE 15
HOW DO YOU TRAIN THE AI?
SLIDE 16
TRAINING
BMW X5 Ford Ecosport Hyundai i30
SLIDE 17
LOTS OF THEM
SLIDE 18
SERIOUSLY LOTS OF THEM
SLIDE 19
10,000,000 CARS
SLIDE 20
10,000,000 CARS LABELLED!
SLIDE 21
SLIDE 22
SLIDE 23
FROM MULTIPLE ANGLES
SLIDE 24
- Huge effort
- Not practical
PROBLEMS
SLIDE 25
MOST WORLD DATAS ARE UNLABELED
- Twitter feeds
- Facebook photos
- Youtube videos
SLIDE 26
UNSUPERVISED LEARNING
SLIDE 27
UNSUPERVISED DEEP LEARNING GENERATING FACES
SLIDE 28
IMAGE GENERATION
- Variational Auto Encoder (VAE)
- Generative Adversarial Network
(GAN)
SLIDE 29
VAE
- Understand the subject (face)
- Generate new subject
SLIDE 30
HOW DOES IT WORK?
SLIDE 31
AUTOENCODER
Encoder Decoder
80x80 40x40 20x20 50x1 20x20 40x40 80x80
Latent Vector
SLIDE 32
- MSE, FPL
Error Minimize
Images
SLIDE 33
LATENT VECTORS
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
SLIDE 34
LATENT VECTORS
Hair-length
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
SLIDE 35
LATENT VECTORS
Hair-length
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
Skin-color
SLIDE 36
LATENT VECTORS
Skin-color Hair-length Eye-size Gender Age
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
SLIDE 37
REMEMBER THE SIGNIFICANT DIFFERENCES HOW COULD IT BE POSSIBLE? SMALL LATENT DIMENSION
SLIDE 38
PACKING A TRAVEL BAG
SLIDE 39
BIG LUGGAGE
SLIDE 40
SMALL BAG
Latent Vector Important Features Learned
SLIDE 41
Info 1 Info 2 Info 3 Info 4 Info 1 Info 2 Info 3 Info 4 Gender Hair Color Skin Color Age
MEMORY GAMES
SLIDE 42
HOW DO WE USE IT TO GENERATE NEW FACE?
SLIDE 43
AUTOENCODER
Encoder Decoder
80x80 40x40 20x20 50x1 20x20 40x40 80x80
Latent Vector
SLIDE 44
AUTOENCODER
Decoder
50x1
Latent Vector
20.1, 10.5, -5.2, 6.2, …
SLIDE 45
AUTOENCODER
Decoder
50x1
Latent Vector
20.1, 10.5, -5.2, 6.2, …
SLIDE 46
LATENT SPACE DISTRIBUTION
Land of no face
SLIDE 47
LATENT SPACE
SLIDE 48
MORPHING
SLIDE 49
FACE ARITHMETIC
SLIDE 50
LATENT VECTORS
Skin-color Hair-length Eye-size Gender Age
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
SLIDE 51
LATENT VECTORS
Hair-length
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
SLIDE 52
LATENT VECTORS
Hair-length
8.5 5.2 8.7
- 3.2
1.5
- 2.4
4.3 4.5 5.4 4.2
- 2.0
0.9 4.3
- 3.5
1.4
Original Image Reconstructed Image
0.7 0.1 0.2
SLIDE 53
20.1, 10.5, -5.2 5.0, 6.3, -5.6 15.1, 4.2, 0.4
Bang vector
SLIDE 54
BANG PREVIEW
1.0, 5.3, 3.2 15.1, 4.2, 0.4
Bang vector
16.1, 9.5, 3.6
SLIDE 55
ADDING GLASSES
SLIDE 56
REMOVING GLASSES
SLIDE 57
ADDING YOUTH
SLIDE 58
GAN GENERATIVE ADVERSARIAL NETWORK
SLIDE 59
Real Face Images Discriminator Network Random Vector Generative Network Generated Face Image Prediction Real or Fake Punished on Discriminator Network’s success Punished on Discriminator Network’s failure
SLIDE 60
MORPHING USING GAN
SLIDE 61
SLIDE 62
OTHER APPLICATION OF GENERATIVE NETWORK
SLIDE 63
SUPER RESOLUTION
SLIDE 64
SUPER RESOLUTION
SLIDE 65
SUPER RESOLUTION
SLIDE 66
SUPER RESOLUTION
SLIDE 67
SUPER RESOLUTION
SLIDE 68
SUPER RESOLUTION
SLIDE 69
SUPER RESOLUTION
64x128 256x512 4x
SLIDE 70
SUPER RESOLUTION
SLIDE 71
WHY IS IT IMPORTANT? UNSUPERVISED LEARNING
SLIDE 72
AGI
SLIDE 73
WHAT’S NEEDED FOR AN AGI
SLIDE 74
SELF LEARNING
Supervised Unsupervised
SLIDE 75
KNOWLEDGE EXTRACTION
SLIDE 76
REASONING
Knowledge Extraction Information Reasoning
SLIDE 77
WITHOUT REASONING
SLIDE 78
WITHOUT REASONING
SLIDE 79
TAKE AWAY NOTHING
SLIDE 80
DANGER !!!
SLIDE 81
THANK YOU
SLIDE 82