A path to unsupervised learning

Soumith ChintalaFacebook AI Research

through Adversarial Networks

Overviewof the talk

• Unsupervised Learning • Generative Adversarial Networks • Advances • Using the learnt representations • What’s next?

Unsupervised LearningAn introduction

Unsupervised LearningAn introduction

Supervised Learning

Unsupervised LearningAn introduction

Unsupervised Learning

Unsupervised LearningUsefulness

Unsupervised LearningReusing representations

Generative ModelsAn introduction

A model that learns a distribution of images

Generative ModelsAn introduction

X = P(z), z controls dogness or catness

Generative ModelsAn introduction

X = P(z), z is a latent variable

Generative ModelsAn introduction

P(z) = neural network

Generative Adversarial Networks

Generative Adversarial NetworksAlternating optimization

Generator Sample Optimizer

Training Data

Loss: Looks Real

Generative Adversarial Networks

Generative Adversarial NetworksAlternating optimization

Generatornoise SampleClassification

Loss

Training Data

Learnt Real/Fake Cost function

Discriminator

Generative Adversarial NetworksAlternating optimization

Generatornoise SampleClassification

Loss

Training Data

Discriminator

Trained via Gradient Descent

Generative Adversarial NetworksAlternating optimization

Generatornoise SampleClassification

Loss

Training Data

Discriminator

Optimizing to fool D

Generative Adversarial NetworksAlternating optimization

Generatornoise SampleClassification

Loss

Training Data

Discriminator

Optimizing to not get fooled by G

Generative Adversarial NetworksOptimizes Jensen-Shannon Divergence

Generative Adversarial NetworksSamples

Class-conditional GANs

Generatornoise

SampleClassification

Loss

Training Data

Discriminator

Class-conditional GANsNot unsupervised

class

Video Prediction GANs

Generatornoise SampleClassification

Loss

Training Data

Discriminator

Video Prediction GANs

Generatornoise SampleClassification

Loss

Training Data

Discriminator

Video Prediction GANs

Generatornoise SampleClassification

Loss

Training Data

Discriminator

MSE Loss

Video Prediction GANs

DCGANs

Latent space arithmetic

Using the GAN feature representation

Using the GAN feature representation

Using the GAN feature representation

Needs much lesser labeled data

Using the GAN feature representation

In-painting GANs

In-painting GANs

In-painting GANs

Disentangling representations

Disentangling representations

Disentangling representations

Disentangling representations

Disentangling representations

Disentangling representations

Stability and Representation Reuse

Stability and Representation Reuse

• Feature matching • Minibatch discrimination • Label smoothing • What’s next?

Stability and Representation Reuse

Stability and Representation Reuse

What’s next?• Planning and forward modeling

Questions

• When will adversarial networks take over the world? • Soon.