CHAPTERS
1.
Welcome back to ICML 2019 presentations. This session on Generative Models includes:
00:00
2.
[Paper: Tensor Variable Elimination for Plated Factor Graphs]
05:25
3.
Learning and inference with discrete latent variables
05:50
4.
Background: Factor graph inference
07:10
5.
Focus: Plated factor graphs
08:54
6.
Plated factor graph inference
09:48
7.
Challenges: Plated factor graph inference
10:29
8.
Algorithm: Tensor variable elimination
11:21
9.
Algorithm: Computational complexity
15:49
10.
Implementation: exploiting existing software
17:35
11.
Implementation: Integration with the Pyro PPL
18:13
12.
Implementation: Scaling with parallel hardware
18:52
13.
Experiments
19:51
14.
Experiment 1: Polyphonic Music Modeling
20:21
15.
Experiment 2: Animal population movement
21:14
16.
Experiment 3: word sentiment from weak supervision
22:00
17.
Q&A
23:23
18.
[Paper: Predicate Exchange]
25:26
19.
Objective
25:41
20.
Priors with constraints
26:35
21.
Inverse Graphics with constraints
27:06
22.
Predicate Exchange
27:42
23.
Convert predicates into soft predicates
28:28
24.
Approximate Posterior
28:55
25.
Temperature trades-off accuracy / convergence
29:17
26.
Replica Exchange
29:46
27.
Omega.jl: A Causal, Higher-Order PPL
30:07
28.
[Paper: Discriminative Regularization for Latent Variable Models with Applications to Electrocardiography]
30:54
29.
Black-Box Predictors
31:03
30.
Gradient-based Explanations
31:42
31.
Generative Latent Variable Models
32:16
32.
Unsupervised Models Drawbacks
32:57
33.
Discriminatively Regularized VAE (DR-VAE)
33:47
34.
EKGs: Discriminative-Generative Tradeoff
34:21
35.
Model-Morphs
34:51
36.
More details at the poster!
35:11
37.
[Paper: Hierarchical Decompositional Mixtures of Variational Autoencoders]
35:38
38.
In Short
35:53
39.
Sum Product Network
36:51
40.
Variational Autoencoder
38:12
41.
Sum-Product Variational Autoencoder
39:16
42.
Empirical Results
39:53
43.
Thank You
40:20
44.
[Paper: Finding Mixed Nash Equilibria of Generative Adversarial Networks]
40:44
45.
Learning distributions
40:57
46.
Wasserstein GANs
41:14
47.
Wasserstein GANs: From pure to mixed Nash Equilibrium
42:34
48.
A re-thinking of GANs via the mixed Nash equilibrium
43:30
49.
Entropic Mirror Descent Iterates in Infinite Dimension
44:17
50.
A Practical Algorithm
44:27
51.
Thanks!
45:04
52.
[Paper: CompILE: Compositional Imitation Learning and Execution]
45:39
53.
Skill Discovery from Demonstrations
46:03
54.
Compositional Imitation Learning and Execution
47:29
55.
Imitation Learning
48:29
56.
ComplLE inference model
50:47
57.
CompILE generative model
52:11
58.
Soft sequence masking
52:54
59.
ComplLE for Imitation Learning
53:14
60.
Qualitative results
55:03
61.
Quantitative results - Imitation Learning
56:08
62.
Continuous state/action spaces
58:59
63.
Using the learned sub-policies
1:01:23
64.
CompILE - Summary and conclusions
1:03:01
65.
Q&A
1:03:43
66.
Welcome back to ICML 2019 presentationfacebook research This session on Generative Models includes:
1:04:46
67.
[Paper: Sparse Multi-Channel Variational Autoencoder for the Joint Analysis of Heterogenous Data]
1:05:24
68.
The Generative Multi-Channel Model
1:05:33
69.
Unsupervised clustering in Alzheimers' Disease
1:08:19
70.
Generation from latent space
1:08:56
71.
Thank you!
1:09:24
72.
[Paper: Learning Deep Generative models via Variational Gradient Flow (VGrow)]
1:09:49
73.
Generative model
1:10:12
74.
Intuition
1:11:08
75.
Variational Gradient Flow (VGrow)
1:11:21
76.
Find h( )
1:11:44
77.
Generated Portrait (based on Wiki Art)
1:13:20
78.
Connection with Differential Equation
1:13:40
79.
Summary
1:14:07
80.
[Paper: Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design]
1:15:00
81.
Overview
1:15:52
82.
Continuous flows for discrete data
1:17:21
83.
Variational Dequantization
1:18:48
84.
Coupling layers
1:19:06
85.
Ablation on CIFAR
1:19:22
86.
[Paper: Learning Neurosymbolic Generative Models via Program Synthesis]
1:20:14
87.
Generative Models and Global Structure
1:20:41
88.
Our Approach: Global Structure as Programs
1:21:44
89.
Application to Image Completion
1:22:17
90.
Learning via Program Synthesis - Phase 1
1:22:46
91.
Learning via Program Synthesis - Phase 2
1:23:36
92.
Comparison to Baselines
1:23:57
93.
Experimental Results
1:24:25
94.
Future Work
1:25:00
Powered by
VideoKen
MORE
Share
COPIED
Share Topic...
×
Search in Video
Feedback
Powered by
VideoKen
SEARCH IN VIDEO
Transcript is not available for this video
Powered by
VideoKen
FEEDBACK
Love it, will recommend
Good, but won't recommend
Bad, needs a lot of improvement
Submit Feedback
Powered by
VideoKen