Sign inSign up
Deep Generative Models thumbnail
Pause
Mute
Subtitles not available
Playback speed
0.25
0.5
0.75
1
1.25
1.5
1.75
2
Full screen

Deep Generative Models

Published on Sep 13, 201520155 Views

Yoshua Bengio

Deep Learning Summer School 2015 - Montreal

Related categories

Deep LearningReinforcement LearningUnsupervised Learning

Chapter list

Deep Generative Models  00:00
Learning ''How the world ticks''00:28
Learning Multiple Levels of abstraction04:00
Invariance and Disentangling05:20
Emergence of Disentangling07:26
Why Latent Factors & Unsupervised Representation Learning? Because of Causality09:38
Challenge with Graphical Models with Latent Variables14:38
Log-likelihood Gradient in Undirected Graphical Models16:54
Issues with Maximum likelihood for Boltzmann Machines19:44
Poor mixing: Depth to the Rescue23:48
Space-Filling in Representation-Space25:19
Bypassing Normalization Constants with Generative Black Boxes29:41
Denoising Auto-Encoders doing Score Matching on Gaussian RBMs38:35
Score Matching42:18
Denoising Auto-Encoder42:30
Learning a Vector Field that Estimates a Gradient Field43:16
Preference for Locally Constant Features43:17
Denoising Score Matching45:22
Denoising Auto-Encoder Markov Chain45:25
Auto-Encoders46:02
Denoising Auto-Encoders learn a Markov Chain Transition Distribution46:29
Many Modes challenge47:10
Consistency Results48:16
Generative Stochastic Networks49:06
Generative Stochastic Networks (GSN)55:12
GSN Experiments55:53
Not just MNIST55:53
GSNs/DAEs can model complex distributions and missing modalities55:55
Ancestral Sampling with Learned Approximate Inference55:56
Extracting Structure By Gradual Disentangling and Manifold Unfolding 55:57
NICE57:45
NCIE Samples01:00:45
NICE Inpainting01:00:58
NICE Inpainting Movies01:01:43
NICE: Perfect Auto-Encoders01:01:43
Bypassing Normalization Constants with Generative Black Boxes01:05:06
Generative adversarial networks01:05:12
Adversarial nets framework - 101:05:19
Adversarial nets framework - 201:08:13
Zero-sum game01:09:33
Police (Discriminator) vs Counterfeiter (Generator) 01:09:34
Learning process01:09:35
Generated Samples01:12:36
Visualization of model samples01:12:45
Learned 2-D manifold of MINST01:12:46
Visualization of model trajectories - 101:12:48
Visualization of model trajectories - 201:13:40
Laplacian Pyramid of Conditional GANs 01:14:44
LAPGAN results01:17:50
Other Encouraging News: Semisupervised Learning with Ladder Network 01:18:54
Outstanding Results01:22:15
Conclusions01:24:38