Learning Deep Generative Models

Published on 2016-08-2314492 Views

Ruslan Salakhutdinov

In this tutorial I will discuss mathematical basics of many popular deep generative models, including Restricted Boltzmann Machines (RBMs), Deep Boltzmann Machines (DBMs), Helmholtz Machines, Variatio

Deep Learning Summer School 2016 - Montreal

Related categories

Presentation

Learning Deep Generative Models 00:00

Mining for Structure/100:02

Mining for Structure/200:46

Example: Understanding Images01:01

Restricted Boltzmann Machines03:38

Learning Features05:15

Model Learning/1 05:44

Model Learning/207:42

RBMs for Real-valued Data/108:22

RBMs for Real-valued Data/209:04

RBMs for Real-valued Data/309:10

RBMs for Word Counts/109:26

RBMs for Word Counts/210:00

Different Data Modalities 10:36

Product of Experts/111:24

Product of Experts12:52

Deep Boltzmann Machines/112:58

Deep Boltzmann Machines/213:09

Model Formulation13:48

Mathematical Formulation15:00

Approximate Learning/115:48

Approximate Learning/216:04

Approximate Learning17:23

Stochastic Approximation18:05

Learning Algorithm19:10

Variational Inference/121:26

Variational Inference/226:09

Variational Inference/327:13

Variational Inference/428:47

Good Generative Model?/1 28:49

Good Generative Model?/229:15

Good Generative Model?/3 29:22

Good Generative Model?/4 29:30

Good Generative Model?/529:36

Good Generative Model?/630:11

Handwriting Recognition30:40

Generative Model of 3-D Objects30:47

3-D Object Recognition 31:18

Learning Hierarchical Representations 31:38

Talk Roadmap31:51

Helmholtz Machines31:54

Various Deep Generative Models34:12

Motivating Example 35:04

Overall Model 35:55

Flipping Colors 36:34

Flipping Backgrounds 37:06

Flipping Objects 37:31

Qualitative Comparison 38:00

Variational Lower-Bound 38:15

Novel Scene Compositions38:45

Overall Model40:21

Variational Autoencoders (VAEs)41:18

VAE: Example 42:43

Recognition Network43:16

Variational Bound 44:37

Reparameterization Trick/146:43

Reparameterization Trick/247:48

Computing the Gradients/148:49

Computing the Gradients/250:27

VAE: Assumptions51:46

Importance Weighted Autoencoders/152:25

Importance Weighted Autoencoders/254:10

Tighter Lower Bound54:51

Computing the Gradients 56:12

IWAEs vs. VAEs/156:42

IWAE: Intuition/157:53

IWAE: Intuition/258:46

Computation with IWAEs59:27

Two Architectures 01:00:08

MNIST Results/1 01:01:14

IWAEs vs. VAEs/201:01:23

IWAEs vs. VAEs/301:01:27

MNIST Results/2 01:02:02

Key Observation01:02:39

Talk Roadmap01:05:56

Caption Generation/1 01:06:17

Encode-Decode Framework01:06:36

Caption Generation/2 01:06:48

Caption Generation/3 01:07:10

Caption Generation with Visual Attention01:07:31

Visual Attention01:07:57

Improving Action Recognition01:08:18

Recurrent Attention Model/1 01:08:29

Recurrent Attention Model/2 01:09:46

Model Definition01:10:25

Variational Learning/101:11:16

Variational Learning/201:11:59

Variational Learning/301:12:52

Sampling from the Prior01:13:09

Maximizing Marginal Likelihood01:14:20

Comparing the Two Estimators01:14:56

Another Key Observation/101:15:51

Another Key Observation/201:16:29

Relationship To Helmholtz Machines/101:16:39

Relationship To Helmholtz Machines/2 01:16:59

The Wake-Sleep Recurrent Attention Model 01:18:18

Training Inference Network01:18:33

MNIST Attention Demo01:18:58

Hard vs. Sod Attention 01:19:25

Talk Roadmap 01:20:07

(Some) Open Problems/1 01:20:17

(Some) Open Problems/2 01:20:28

Sequence to Sequence Learning 01:21:14

Skip-Thought Model/1 01:21:26

Skip-Thought Model/2 01:21:40

Learning Objective01:21:52

Semantic Relatedness01:22:06

Semantic Relatedness Recurrent Neural Network 01:22:31

Neural Story Telling 01:24:51

Hierarchical RNNs/101:25:37

Hierarchical RNNs/201:25:53

Atari Games01:26:20

Actor-Mimic Net in Action01:26:39

Transfer Learning01:26:42

Summary01:26:46

Thank You01:27:35