Predicting the Future using Deep Adversarial Networks
Soumith ChintalaFacebook AI Research
Learning With No Labeled Data
Overviewof the talk
• The problem at hand • What are the benefits? • How did we solve it • What have we achieved • What’s left?
You are here
You are here
Walk here!
Route 1
Route 2
Route 3
Let’s Train a Route Generator
Route Generator
Training Data
Linear Regressor with Mean-square Error
Route Generator
RouteGenerator
Linear Regressor with Mean-square Error
Route Generator
RouteGeneratornoise
Linear Regressor with Mean-square Error
Route Generator
RouteGeneratornoise Route
Linear Regressor with Mean-square Error
Route Generator
RouteGeneratornoise Route Optimizer
Training Data
Loss: MSE
Route Generator
Training Data
Linear Regressor with Mean-square Error
Route Generator
Eh?
Linear Regressor with Mean-square Error
Route Generator
Eh?
Linear Regressor with Mean-square Error
Route Generator
Eh? Eh?
Linear Regressor with Mean-square Error
Route Generator
Eh? Eh? Eh?
Linear Regressor with Mean-square Error
Route Generator
Eh? Eh? Eh? Eh?
Linear Regressor with Mean-square Error
Route Generator
Eh? Eh? Eh? Eh? Eh?
Linear Regressor with Mean-square Error
Route Generator
Problem!Converges to the mean
of training samples
Linear Regressor with Mean-square Error
Route Generator
Which is not a valid route!
Problem!Converges to the mean
of training samples
Linear Regressor with Mean-square Error
Route Generator
Let’s try again!
Linear Regressor with Route Validator
Route Generator
RouteGenerator
noiseRoute Optimizer
Training Data
Loss: MSEpoints
Route Generator
RouteGenerator Route Optimizer
Training Data
Loss: Valid Route
Linear Regressor with Route Validator
noise
points
Route Generator
RouteGenerator Route Optimizer
Training Data
Loss: Valid Route
Linear Regressor with Route Validator
noise
points
Route Generator
RouteGenerator Route Optimizer
Training Data
Loss: Valid Route
Linear Regressor with Route Validator
????
noise
points
Generator
Generator Sample Optimizer
Training Data
Loss: Looks Real????
noise
points
Generator
Generatornoise Sample Optimizer
Training Data
Loss: Looks Real
Generator
Generatornoise SampleClassification
Loss
Training Data
Learnt Real/Fake Cost function
Discriminator
Generator
Generatornoise SampleClassification
Loss
Training Data
Neural Net
Discriminator
Neural Net
Generator
Generatornoise SampleClassification
Loss
Training Data
Discriminator
Trained via Gradient Descent
Generator
Generatornoise SampleClassification
Loss
Training Data
Discriminator
Optimizing to fool D
Generator
Generatornoise SampleClassification
Loss
Training Data
Discriminator
Optimizing to not get fooled by G
Route 4
Route Crazy!
Route Crazy! But valid.
Generator
Generatornoise
SampleClassification
Loss
Training Data
Discriminator
Optimizing to not get fooled by G
class
Generator
Generatornoise SampleClassification
Loss
Training Data
Discriminator
Optimizing to not get fooled by G
Generator
Generatornoise SampleClassification
Loss
Training Data
Discriminator
Optimizing to not get fooled by G
MSE Loss
Uses
• Unsupervised Learning • Learn when there’s little labeled data
• Planning • Look-ahead to take better decisions