finish 2nd half

GRAM-nets · Apr 8, 2020 · 5c8644f · 5c8644f
1 parent 0085295
commit 5c8644f
Show file tree

Hide file tree

Showing 11 changed files with 35 additions and 11 deletions.
diff --git a/.gitignore b/.gitignore
@@ -1 +1,2 @@
+.DS_Store
 *.code-workspace
diff --git a/comp_graph-gan.png b/comp_graph-gan.png
diff --git a/comp_graph-gramnet.png b/comp_graph-gramnet.png
diff --git a/comp_graph-mmdgan.png b/comp_graph-mmdgan.png
diff --git a/comp_graph-mmdnet.png b/comp_graph-mmdnet.png
diff --git a/gram-comp-graph.png b/gram-comp-graph.png
diff --git a/iclr2020.md b/iclr2020.md
@@ -36,7 +36,7 @@ Adversarial Generative Models(GANs, MMD-GANs)
 
 ---
 
-## GRaM: Algorithm
+## GRaM: the algorithm
 
 :one: Learn the manifold projection function $f_\theta(x)$ by minimising the squared difference between the pair of density ratios:
 
@@ -52,6 +52,8 @@ $$
 
 ---
 
+## GRaM: the algorithm (continued)
+
 :two: Train the generator $G\gamma$ by minimizing the empirical estimator of MMD in the low-dimensional manifold,
 
 $$
@@ -64,7 +66,9 @@ $$
 
 ---
 
-## Pearson Divergence Maximization and Density Ratio Estimation
+## GRaM: the algorithm (continued)
+
+### Pearson divergence maximization and density ratio estimation
 
 Monte Carlo approximation of PD,
 $$
@@ -82,38 +86,57 @@ We use a MMD based density ratio estimator (Sugiyama et al., 2012) under the fix
 
 ---
 
-## The Ring dataset: Illustration of the Method and Stability
+## How do GRAM-nets compare to other methods
 
-![Ring: training process](syn_0.png)
+| GAN | MMD-net | MMD-GAN | GRAM-net |
+| - | - | - | - |
+| ![Computation graph (GAN)](comp_graph-gan.png) | ![Computation graph (MMD-net)](comp_graph-mmdnet.png) | ![Computation graph (MMD-GAN)](comp_graph-mmdgan.png) | ![Computation graph (GRAM-net)](comp_graph-gramnet.png) |
 
 ---
 
-![Ring: stability on MMD-nets and MMD-GANs](syn_1.png)
+## Illustration of GRAM training
+
+| | |
+| - | - |
+| ![Ring: training process](ring_process.png) | Blue: data<br> Orange: samples<br><br> Top: original<br> Bottom: projected|
 
 ---
 
-![Ring: stability on GAN and GRAM-nets](syn_2.png)
+## Evaluations: the stability of models
+
+![Ring: stability on GAN and GRAM-nets](stability_gan_gramnet.png)
+
+x-axis = noise dimension and y-axis = generator layer size
 
 ---
 
-## Quantitative Results: Sample Quality
+## Evaluations: the stability of models (continued)
+
+![Ring: stability on MMD-nets and MMD-GANs](stability_mmdnet_mmdgan.png)
+
+x-axis = noise dimension and y-axis = generator layer size
+
+---
+
+## Quantitative results: sample quality
 
 ![FID table](table_image.png)
 
 ---
-## Qualitative Results: Random Samples
+
+## Qualitative results: random samples
 
 ![Samples (CIFAR10)](cifar10.png) ![Samples (CelebA)](celeba.png)
 
 ---
 
-## The End
+## The end
 
 Extra slides to follow...
 
 ---
 
-## Density Ratio Estimation via (Infinite) Moment Matching
+## Density ratio estimation via (infinite) moment matching
 
 ### Maximum mean discrepancy
 
@@ -136,7 +159,7 @@ $$
 
 ---
 
-## Generator Training
+## Generator training
 
 The generator $G\gamma$ is trained by minimizing the empirical estimator of MMD
 

diff --git a/ring_process.png b/ring_process.png
diff --git a/syn_2.png → stability_gan_gramnet.png b/syn_2.png → stability_gan_gramnet.png
diff --git a/syn_1.png → stability_mmdnet_mmdgan.png b/syn_1.png → stability_mmdnet_mmdgan.png
diff --git a/syn_0.png b/syn_0.png