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
SUPERVISED LEARNING IMAGE RECOGNITION
CAR RECOGNITION
Ford Kuga Titanium
SUPERVISED LEARNING I am awesome
HOW DO YOU TRAIN THE AI?
TRAINING BMW X5 Ford Ecosport Hyundai i30
LOTS OF THEM
SERIOUSLY LOTS OF THEM
10,000,000 CARS
10,000,000 CARS LABELLED!
FROM MULTIPLE ANGLES
PROBLEMS • Huge effort • Not practical
MOST WORLD DATAS ARE UNLABELED • Facebook photos • Youtube videos • Twitter feeds
UNSUPERVISED LEARNING
UNSUPERVISED DEEP LEARNING GENERATING FACES
IMAGE GENERATION • Variational Auto Encoder (VAE) • Generative Adversarial Network (GAN)
VAE • Understand the subject (face) • Generate new subject
HOW DOES IT WORK?
AUTOENCODER 20x20 40x40 80x80 80x80 40x40 20x20 50x1 Latent Vector Decoder Encoder
MSE, FPL - Error Images Minimize
LATENT VECTORS Reconstructed Original Image Image 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
LATENT VECTORS Reconstructed Original Image Image 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 Hair-length
LATENT VECTORS Reconstructed Original Image Image 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 Hair-length Skin-color
LATENT VECTORS Reconstructed Original Image Image 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 Hair-length Skin-color Eye-size Gender Age
HOW COULD IT BE POSSIBLE? SMALL LATENT DIMENSION REMEMBER THE SIGNIFICANT DIFFERENCES
PACKING A TRAVEL BAG
BIG LUGGAGE
SMALL BAG Latent Vector Important Features Learned
MEMORY GAMES Info 1 Info 2 Info 3 Info 4 Info 1 Info 2 Info 3 Info 4 Gender Hair Color Skin Color Age
HOW DO WE USE IT TO GENERATE NEW FACE?
AUTOENCODER 20x20 40x40 80x80 80x80 40x40 20x20 50x1 Latent Vector Decoder Encoder
AUTOENCODER 50x1 20.1, 10.5, -5.2, 6.2, … Latent Vector Decoder
AUTOENCODER 50x1 20.1, 10.5, -5.2, 6.2, … Latent Vector Decoder
LATENT SPACE DISTRIBUTION Land of no face
LATENT SPACE
MORPHING
FACE ARITHMETIC
LATENT VECTORS Reconstructed Original Image Image 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 Hair-length Skin-color Eye-size Gender Age
LATENT VECTORS Reconstructed Original Image Image 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 Hair-length
LATENT VECTORS Reconstructed Original Image Image 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 0.7 0.1 0.2 Hair-length
20.1, 10.5, -5.2 5.0, 6.3, -5.6 Bang vector 15.1, 4.2, 0.4
BANG PREVIEW Bang vector 1.0, 5.3, 3.2 15.1, 4.2, 0.4 16.1, 9.5, 3.6
ADDING GLASSES
REMOVING GLASSES
ADDING YOUTH
GAN GENERATIVE ADVERSARIAL NETWORK
Real Face Discriminator Prediction Images Network Real or Fake Punished on Discriminator Network’s Generated failure Face Image Punished on Discriminator Network’s Generative Random Vector success Network
MORPHING USING GAN
OTHER APPLICATION OF GENERATIVE NETWORK
SUPER RESOLUTION
SUPER RESOLUTION
SUPER RESOLUTION
SUPER RESOLUTION
SUPER RESOLUTION
SUPER RESOLUTION
SUPER RESOLUTION 256x512 64x128 4x
SUPER RESOLUTION
UNSUPERVISED LEARNING WHY IS IT IMPORTANT ?
AGI
WHAT’S NEEDED FOR AN AGI
SELF LEARNING Supervised Unsupervised
KNOWLEDGE EXTRACTION
REASONING Reasoning Knowledge Extraction Information
WITHOUT REASONING
WITHOUT REASONING
TAKE AWAY NOTHING
DANGER !!!
THANK YOU
QUESTION?
Recommend
More recommend