-
-
Notifications
You must be signed in to change notification settings - Fork 0
/
slides_7_orig.html
612 lines (522 loc) · 26.3 KB
/
slides_7_orig.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
<!doctype html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<meta name="author" content="Emre Neftci">
<title>Neural Networks and Machine Learning</title>
<link rel="stylesheet" href="css/reset.css">
<link rel="stylesheet" href="css/reveal.css">
<link rel="stylesheet" href="css/theme/nmilab.css">
<link rel="stylesheet" type="text/css" href="https://cdn.rawgit.com/dreampulse/computer-modern-web-font/master/fonts.css">
<link rel="stylesheet" type="text/css" href="https://maxcdn.bootstrapcdn.com/font-awesome/4.5.0/css/font-awesome.min.css">
<link rel="stylesheet" type="text/css" href="lib/css/monokai.css">
<!-- Printing and PDF exports -->
<script>
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match( /print-pdf/gi ) ? 'css/print/pdf.css' : 'css/print/paper.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
<script async defer src="https://buttons.github.io/buttons.js"></script>
</head>
<body>
<div class="reveal">
<div class="slides">
<!--
<section data-markdown><textarea data-template>
<h2> </h2>
<ul>
<li/>
</ul>
</textarea></section>
[](https://colab.research.google.com/github/surrogate-gradient-learning/pytorch-lif-autograd/blob/master/tutorial01_dcll_localerrors.ipynb)
-->
<section data-markdown data-vertical-align-top data-background-color=#B2BA67><textarea data-template>
<h1> Neural Networks and Machine Learning </h1>
<h2> Week 7: Deep Generative Models</h2>
### Instructors: Emre Neftci and Takashi Nagata
<center>https://canvas.eee.uci.edu/courses/21750</center>
<center>http://tinyurl.com/nmi-lab-appointments</center>
[](?print-pdf)
</textarea>
</section>
<section data-markdown data-vertical-align-top><textarea data-template>
<h2> Goody Bag IV</h2>
<ol>
<li/> Data Augmentation
<li/> t-Distributed Stochastic Neighbor Embedding (t-SNE)
</ol>
</textarea>
</section>
<section data-markdown><textarea data-template>
<h2> Data Augmentation</h2>
<ul>
<li/> Data augmentation is a key trick used to improve the test performance of deep neural networks, especially when the dataset is small.
<li/> Increasing the datset size has an effect similar to regularization, by preventing overfitting.
<li/> Several methods exist to augment the data:
<ol>
<li/> Random Crop
<li/> Random flips
<li/> Color jitter
<li/> ...
</ol>
<li/> Pytorch torchvision.transforms has a long list of random transformations
<li/> You can be creative and chain multiple transforms using transforms.Compose
</ul>
[](https://drive.google.com/open?id=1_9KnC7ezIre-n6G_6uC0oUHJzgxoRUyZ)
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Visualization by Dimensionality Reduction </h2>
<ul>
<li/> PCA captures the directions of maximum variance. By limiting the number of directions to 2 or 3, we can visualize the dataset.
<img src="images/pca.png" class=small />
<li /> But PCA is just a linear projection, and the data many still be hard to visualize.
<li /> A better visualization would be a dimensionality reduction that preserves the distances between points, because this will preserve the geometry of the data.
<ul>
<li /> Assume $d_{ij}$ is the original distance between point i and j, and $\hat{d}_{ij}$ is the distance in the projected space. We would like to minimize
$$ C = \sum_{i \ne j} (d_{ij}-\hat{d}_{ij})^2 $$
<li /> t-SNE is the most popular method for dimensionality reduction in deep learning.
</ul>
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Goody IV.1: t-Distributed Stochastic Neighbor Embedding (t-SNE) </h2>
<ul>
<li/> t-SNE builds a neighbor relation ship between points and strives to keep the number of neighbors the same when projecting into the reduced space.
<li /> However it is an optimization routine in itself which can get stuck in local minima (more data the better!)
<img src="images/t_sne_example.png"/>
<pre><code class="Python" data-trim data-noescape>
import numpy as np
from sklearn.manifold import TSNE
lat_tsne = TSNE(n_components=2).fit_transform(lat_z)
</code></pre>
<a href=http://jmlr.csail.mit.edu/papers/volume9/vandermaaten08a/vandermaaten08a.pdf><p class=ref>Van det Maaten et al. 2008</p></a>
</ul>
</textarea></section>
<!-- .element: class="fragment" -->
<section data-markdown data-vertical-align-top><textarea data-template>
<h2> Generative Models </h2>
<ul>
<li /> In high dimensional spaces, many vectors $\mathbf{x}$ result in $p(\mathbf{x})$ that vanish. It is more useful to focus around points that are likely.
<li /> For example, if we wish to estimate the half of a digit, it would be very helpful to first decide which digit to draw. The digit forms the <em> latent variable </em> (code), which we called $\mathbf{z}$.
<img src="images/Samples-drawn-from-the-prediction-of-the-lower-half-of-the-MNIST-test-data-digits-based.png" class=small />
<li /> For this, we need to create a relationshop between $\mathbf{z}$ and $\mathbf{x}$. This means we aim to maximize $p(x)$ for every $\mathbf{x}$ in the dataset, <em> i.e. </em>
$$ p(\mathbf{x}) = \int_\mathbf{z} p(\mathbf{x}|\mathbf{z}) p(\mathbf{z}) \mathrm{d}\mathbf{z} $$
where $\mathbf{z}$ is a latent variable
<li /> There are two challenges: how to decide on $\mathbf{z}$, and how to perform the integral.
</ul>
</textarea>
</section>
<section data-markdown><textarea data-template>
<h2> Variational AE (VAE) </h2>
<ul>
<li /> Variational Autoencoders were proposed as approximate but fast algorithms to maximizing $p(\mathbf{x})$
<li /> VAEs assert that $\mathbf{z}$ are drawn from a simple distribution, such as a multivariate Gaussian $N(0,\mathbb{I})$.
<li /> How can this $\mathbf{z}$ be powerful enough? Given a random variable with one distribution, it is possible to create any other distribution provided a complex enough function.
<img src="images/from_gaussian_to_arbitrary.png">
<p class=ref>Doersch, 2016, <em>Tutorial on Variational Autoencoders</em> </p>
<li /> So from independent Gaussian-distributed variables, we can obtain any latent variable distribution, which can then be mapped to $\mathbf{x}$ using some function $f$.
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Variational AE (VAE) </h2>
<ul>
<li/> VAEs use a Gaussian distribution in $p(\mathbf{x}|\mathbf{z})$, with mean $f_\theta(z)$ and variance $\sigma^2 \mathbb{I}$
$$ p(\mathbf{x}) = \int_\mathbf{z} N(\mathbf{x}|f_\theta(\mathbf{z}), \sigma^2 \mathbb{I}) p(\mathbf{z}) \mathrm{d}\mathbf{z} = \mathbb{E}_{z\sim p} N(\mathbf{x}|f_\theta(\mathbf{z}), \sigma^2 \mathbb{I})$$
(but any density that can be computed and differentiable with respect to the parameters $\theta$ can be used)
<li /> Unfortunately, if $p(\mathbf{z})$ is also a Gaussian, many samples $\mathbf{z}$ are required to estimate $p(\mathbf{x})$.
<li />VAEs solve this problem by restricting the integral to values of $\mathbf{z}$ that are likely to have produced $\mathbf{x}$. This is achieved by using a density $q(\mathbf{z}|\mathbf{x})$ instead of $p(\mathbf{z})$.
<!-- We now have:
$$ p(\mathbf{x}) = \mathbb{E}_{\mathbf{z}\sim q} N(\mathbf{x}|f_\theta(\mathbf{z}), \sigma^2 \mathbb{I}) = \mathbb{E}_{\mathbf{z}\sim q} p(x|z)$$ -->
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Variational Lower Bound </h2>
<ul>
<li /> What is the relationship between $\mathbb{E}_{z\sim q} p(\mathbf{x}|\mathbf{z})$ and $p(\mathbf{x})$?
<li /> The following relationship holds, known as the Variational lower bound.
$$
\log p(\mathbf{x}) \ge \mathbb{E}_{\mathbf{z}\sim q} \log p(\mathbf{x}|\mathbf{z}) - KL(q(\mathbf{z}|\mathbf{x})||p(\mathbf{z}))
$$
<li /> It is the cornerstone of all Variational Bayes techniques.
<li /> By maximizing the right hand side, we optimize $\log p(\mathbf{x})$
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Optimizing the Variational Lower Bound </h2>
$$
\text{ Variational Lower Bound: } \mathbb{E}_{\mathbf{z}\sim q} \log p(\mathbf{x}|\mathbf{z}) - KL(q(\mathbf{z}||\mathbf{x})|p(\mathbf{z}))
$$
<ul>
<li /> Right term: we need to choose $q(\mathbf{z}|\mathbf{x})$. Typically, it is a Gaussian distribution $N(\mathbf{z}| \mu(\mathbf{x},\theta), \Sigma(\mathbf{x},\theta))$, where $\mu$ and $\Sigma$ are implemented using neural networks. With this choice, the right term can be computed in closed form.
<li /> Left term: We could evaluate it by sampling $\mathbf{z}$ and averaging them. But this is expensive. The solution is to take a "single sample average".
<li /> We now have:
$$
\log p(\mathbf{x}|\mathbf{z}) - KL(q(\mathbf{z}||\mathbf{x})|p(\mathbf{z}))
$$
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Reparametrization Trick</h2>
$$
\log p(\mathbf{x}|\mathbf{z}) - KL(q(\mathbf{z}||\mathbf{x})|p(\mathbf{z}))
$$
<ul>
<li /> But there is a problem: we also need to optimize $q$, which was used to <em>sample </em> $\mathbf{z}$. Stochastic gradient descent cannot deal with stochastic variables (stochastic samples are not differentiable). The trick is to sample $\mathbf{z}$ as follows:
$$ z = \mu(\mathbf{x}) + \Sigma^{\frac12}(\mathbf{x})\odot \omega $$
where $\omega \sim N(0,\mathbb{I})$. This trick only works for certain distributions!
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Variational AE (VAE) Summary </h2>
<ul>
<li /> Variational Autoencoders were proposed as approximate but fast algorithms to maximizing $p(\mathbf{x})$. With generative adversarial networks, they are the current state-of-the-art for generative modeling.
<li /> As in AEs, VAEs consist in an encoder and a decoder function, both of which are neural networks.
<li /> The code $\mathbf{z}$ is a random variable.
<ul>
<li /> Encoder $q_\phi(\mathbf{z}|\mathbf{x})$ is typically a gaussian probability density (dashed).
<li /> Decoder $p_\theta(\mathbf{x}|\mathbf{z})$ (solid) is a likelihood function of choice (Bernouilli, Multinomial, Gaussian etc.)
<img src="images/vae_graph.png" />
</ul>
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Variational AE (VAE) Architecture </h2>
<ul>
<li /> $p_\theta(\mathbf{x}|\mathbf{z})$ is a neural network taking inputs $\mathbf{z}$ and producing an image
<li /> $q_\phi(\mathbf{z}|\mathbf{x})$ is a neural network producing $\mu(\mathbf{x}) $ and $\mathbf{\Sigma}(\mathbf{x})$
<li /> $\mathbf{z}$ is created using the reparamerization trick $\mathbf{z} = \mu(\mathbf{x}) + \Sigma^{\frac12}(\mathbf{x})\odot \omega $
<img src="https://miro.medium.com/max/3374/1*22cSCfmktNIwH5m__u2ffA.png" />
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Variational AE (VAE) Implementation </h2>
<ul>
<li /> The loss function to evaluate is: $$ \log p(\mathbf{x}|\mathbf{z}) - KL(q(\mathbf{z}||\mathbf{x})|p(\mathbf{z})) $$
<ul>
<li /> $ KL(q(\mathbf{z}||\mathbf{x})|p(\mathbf{z})) = KL(N(\mathbf{z}| \mu(\mathbf{x},\theta), \Sigma(\mathbf{x},\theta))||N(0,\mathbb{I}))$. The KL divergence between two Gaussians can be computed analytically.
<li /> $\log p(\mathbf{x}|\mathbf{z})$, i.e. what is the likelihood of the true data $\mathbf{x}$ given $\mathbf{z}$? This depends on the likelihood function $p(\mathbf{x}|\mathbf{z})$
<ul>
<li />Binary Data $\sim$ Bernouilli distribution: Binary cross entropy loss.
<li />Categorical Data $\sim$ Multinomial distribution: Cross entropy loss.
<li />Real Data $\sim$ Gaussian Density: Mean-Squared Error loss.
</ul>
</ul>
</ul>
[](https://drive.google.com/open?id=1Yh1qtakIrsukG_CYr86obSQiLhD3Gw9K)
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Conditional Variational AE (VAE) Implementation </h2>
<ul>
<li /> As in classical autoencoders, by sampling randomly we have no control over which image class is generated (assuming a class type makes sense)
<li /> There are two easy solutions
<ol>
<li /> Identify a subspace of $\mathbf{z}$ corresponding to the digits, and sample in that region
<li /> Augment the encoder input with a class (= concatenate).
<pre><code class="Python" data-trim data-noescape>
torch.cat([data, labels1h], dim = 1)
</code></pre>
</ol>
<li /> In class assignement, use either method to generate digits of the same class.
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Generative Modeling </h2>
<ul>
<li/> Takes a training data consist of samples drawn from a data-generating distribution $p_{data}$
<li/> Learns to represent an estimate of that distribution somehow
<li/> Some of them estimate the model $p_{model}$ explicitly and the others estimate implicitly
<ul>
<li/> Density estimation <img src="images/GenerativeModeling_concept.png" class=stretch />
<p class=ref> NIPS 2016 Tutorial: Generative Adversarial Networks, Ian Goodfellow </p>
<li/> Sample generation (GANs is in this category)
<ul>
<li/> Some examples later
</ul>
</ul>
</ul>
</textarea></section>
<section data-markdown data-vertical-align-top><textarea data-template>
<h2> Approaches to Generative Modeling </h2>
<img src="images/generative_modeling_approaches.png" class=stretch />
<p class=ref> NIPS 2016 Tutorial: Generative Adversarial Networks, Ian Goodfellow 2017 </p>
</textarea>
</section>
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - examples </h2>
<ul>
<li/> High quality image generation (Karras, et al., 2018)
<li /> Too good to believe? I know...
</ul>
<iframe width="560" height="315" src="https://www.youtube.com/embed/G06dEcZ-QTg" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - examples </h2>
<ul>
<li/> Image-to-Image translation (Isola, et al., 2017)
</ul>
<img src="images/pix2pix.png" class=stretch />
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - a brief intro </h2>
<ul>
<li/> A new framework for estimating generative models via an adversarial process
<li /> Simultaneously train two models:
<ul>
<li/>a generative model G that captures the data distribution
<li/>a discriminative model D that estimates the probability that a sample came from the training data rather than G
</ul>
</ul>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - intuition </h2>
<blockquote><p>The generative model can be thought of as analogous to a team of counterfeiters,
trying to produce fake currency and use it without detection, while the discriminative model is analogous to the police, trying to detect the counterfeit currency. Competition in this game drives both teams to improve their methods until the counterfeits are indistiguishable from the genuine articles.</p>
<footer>—Goodfellow, <cite>Generative Adversarial Networks</cite></footer>
</blockquote>
<img src="images/GAN_overview.png" class=stretch />
<p class=ref> Figure credit: An intuitive introduction to Generative Adversarial Networks (GANs)/freecodecamp </p>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - Some background </h2>
<ul>
<li/> Zero-sum game
<li/> Minimax
<ul>
<li/> You maximize your payoff
<li/> Your opponent tries to minimize your payoff
</ul>
<li/> Find Nash equilibrium
<ul>
<li/> Each player is assumed to know the equilibrium strategies of the other players, and no player has anything to gain by changing only their own strategy (source: Wikipedia)
</ul>
</ul>
</textarea></section>
<!--
Add general explanation about Generative Modeling itself and bridge the last lecture and GANs
-->
<section data-markdown><textarea data-template>
<h2> Generative Adversarial Networks - Cost function and how it works </h2>
<h4>$min_G max_D V(D,G) = \mathbf{E}_{x \sim p_{data}(x)} \left[ logD(x) \right] + \mathbf{E}_{z \sim p_z(z)} \left[ log(1 - D(G(z))) \right]$</h4>
<ul>
<li/> $D$: the probability that a sample came from the training data rather than G
<li/> First equation of the right hand side
<ul>
<li/> Sample data $x$ from real distribution
<li/> Since $x$ is sampled from real distribution, $D(x)$ should be close to 1
</ul>
<li/> Second equation of the right hand side
<ul>
<li/> Draw a noise $z$ from distribution $P_z$ and generate data $G(z)$
<li/> $G$ want to fool $D$: Want to make D(G(z)) close to 1
</ul>
<ul>
<li/> D wants to maximize (as close to 0 as possible) this quantity. G wants to minimize.
</ul>
</ul>
<img src="images/logarithm.png" class=stretch />
<figcaption>Logarithm curve</figcaption>
</textarea></section>
<!--
<section data-markdown><textarea data-template>
<h2> Pros and Cons </h2>
<ul>
<li/> Pros
<ul>
<li/> Doesn't require density function
</ul>
<li/> Cons
<ul>
<li/> Finding the Nash equilibrium state is a difficult problem sometimes
</ul>
</ul>
</textarea></section>
-->
<section data-markdown><textarea data-template>
<h2> GAN Implementation</h2>
<ul>
<li/> Two phases
<ul>
<li/> Train the discriminator with 1. real images and 2. fake images
<li/> Train the generator by trying to fool the discriminator
</ul>
<li/> There is a quize in the code
<pre><code class="Python" data-trim data-noescape>
""" Quiz: Think which one you should use? """
#validity = torch.ones(batch, 1).to(device)
#validity = torch.zeros(batch, 1).to(device)
</code></pre>
</ul>
<img src="images/GAN_overview.png" class=stretch />
[](https://drive.google.com/open?id=1sTKjsfAJcaTTZ4y_UNlTG7Fn_HN9t77v)
</textarea></section>
<section data-markdown data-vertical-align-top><textarea data-template>
<h2> Conditional GAN (cGAN) </h2>
<ul>
<li/> Although GANs showed successful image generation ability, there was no way to control or specify a certain type of image to generate
<li/> Conditional Generative Adversarial Nets was proposed a few months later
<ul>
<li/> Provide some extra information y, such as class labels, to the generator and the discriminator
<li/> How to conditioning? Simply concatenate the additional information with the input
</ul>
</ul>
<img src="images/ConditionalAdversarialNet.png" class=stretch />
</textarea></section>
<!--
<section data-markdown data-vertical-align-top><textarea data-template>
<h2> Implement cGAN </h2>
<ul>
<li/> The first cGAN simply concat additional information for conditioning
</ul>
<pre><code class="Python" data-trim data-noescape>
class Generator(nn.Module):
def __init__(self, batch, z_dim, out_shape, num_classes):
super(Generator, self).__init__()
self.batch_size = batch
self.z_dim = z_dim
self.out_shape = out_shape
self.num_classes = num_classes
self.fc1 = nn.Linear(z_dim+num_classes, 256) # simple concat
</code></pre>
</textarea></section>
-->
<section data-markdown><textarea data-template>
<h2> In-class Assignment: Conditioning your GAN</h2>
<ul>
<li/> Give an additional information (= one-hot representation of true class label) to both the generator and the discriminator
<li/> A helper function is defined in the notebook to convert label which you can get from dataloader into one-hot vector. Feel free to use it or implement by yourself
</ul>
[](https://drive.google.com/open?id=17ud64BV2RajqussXMGtGpmXd7p8-Oxvs)
</textarea></section>
<section data-markdown><textarea data-template>
<h2> My work related to generative modeling </h2>
<ul>
<li/> Reinforcement Learning with generative modeling
<ul>
<li/> Simulate possible futures for planning in reinforcement learning task
<li/> A generative model plays a role of the agent's internal "world model"
</ul>
<li/> Generate imaginary trajectories and plan ahead in combination with a recurrent neural network for short term memory (the video below is the real game image, not the generated one)
</ul>
<video width="320" height="240" controls>
<source src="images/carracing_gym.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Reinforcement Learning </h2>
<img src="images/Silver_RL.png" class=stretch />
<p class=ref> A slide from David Silver's RL course </p>
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Preliminary results </h2>
<ul>
<li/> left column: reconstructed images from predicted $z$
<li/> Right column: reconstructed images from real $z$
</ul>
<img src="images/CogGan_30epoch.jpg" class=stretch />
</textarea></section>
<section data-markdown><textarea data-template>
<h2> Preliminary results </h2>
<ul>
<li/> The first 4 rows are original input images
<li/> Last 4 rows are reconstructed images in 30 epochs
<li/> Should be able to generate lookahead trajectries
</ul>
<img src="images/CogGAN_prediction.png" class=stretch />
</textarea></section>
<section data-markdown><textarea data-template>
<h2> References </h2>
<ul>
<li /> Autoencoders: <a href=>Deep Learning, Goodfellow et al. 2015</a>
<li /> Variational Autoencoder: <a href=>Deep Learning, Goodfellow et al. 2015</a>
<li /> Variational Autoencoder: <a href=>Doersch et al. 2016</a>
</ul>
<ul>
<li /> Generative Adversarial Networks: <a href="https://arxiv.org/abs/1406.2661">Goodfellow et al., 2014</a>
<li /> Conditional Generative Adversarial Nets: <a href="https://arxiv.org/abs/1411.1784">Mehdi Mirza and Simon Osindero, 2014</a>
<li/> NIPS 2016 Tutorial: Generative Adversarial Networks: <a href="https://arxiv.org/abs/1701.00160">Goodfellow, 2017</a>
<li/> Some examples introduced in this slides
<ul>
<li/> Progressive Growing of GANs for Improved Quality, Stability, and Variation: <a href="https://arxiv.org/abs/1710.10196">Karras et al., 2017</a>
<li/> Image-to-Image Translation with Conditional Adversarial Networks: <a href="https://arxiv.org/abs/1611.07004">Isola et al., 2017</a>
</ul>
</ul>
</textarea></section>
</div>
</div>
<!-- End of slides -->
<script src="../reveal.js/lib/js/head.min.js"></script>
<script src="js/reveal.js"></script>
<script>
Reveal.configure({ pdfMaxPagesPerSlide: 1, hash: true, slideNumber: true})
Reveal.initialize({
mouseWheel: false,
width: 1280,
height: 720,
margin: 0.0,
navigationMode: 'grid',
transition: 'slide',
menu: { // Menu works best with font-awesome installed: sudo apt-get install fonts-font-awesome
themes: false,
transitions: false,
markers: true,
hideMissingTitles: true,
custom: [
{ title: 'Plugins', icon: '<i class="fa fa-external-link-alt"></i>', src: 'toc.html' },
{ title: 'About', icon: '<i class="fa fa-info"></i>', src: 'about.html' }
]
},
math: {
mathjax: 'https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js',
config: 'TeX-AMS_HTML-full', // See http://docs.mathjax.org/en/latest/config-files.html
// pass other options into `MathJax.Hub.Config()`
TeX: { Macros: { Dp: ["\\frac{\\partial #1}{\\partial #2}",2] }}
//TeX: { Macros: { Dp#1#2: },
},
chalkboard: {
src:'slides_2-chalkboard.json',
penWidth : 1.0,
chalkWidth : 1.5,
chalkEffect : .5,
readOnly: false,
toggleChalkboardButton: { left: "80px" },
toggleNotesButton: { left: "130px" },
transition: 100,
theme: "whiteboard",
},
menu : { titleSelector: 'h1', hideMissingTitles: true,},
keyboard: {
67: function() { RevealChalkboard.toggleNotesCanvas() }, // toggle notes canvas when 'c' is pressed
66: function() { RevealChalkboard.toggleChalkboard() }, // toggle chalkboard when 'b' is pressed
46: function() { RevealChalkboard.reset() }, // reset chalkboard data on current slide when 'BACKSPACE' is pressed
68: function() { RevealChalkboard.download() }, // downlad recorded chalkboard drawing when 'd' is pressed
88: function() { RevealChalkboard.colorNext() }, // cycle colors forward when 'x' is pressed
89: function() { RevealChalkboard.colorPrev() }, // cycle colors backward when 'y' is pressed
},
dependencies: [
{ src: '../reveal.js/lib/js/classList.js', condition: function() { return !document.body.classList; } },
{ src: 'plugin/markdown/marked.js' },
{ src: 'plugin/markdown/markdown.js' },
{ src: 'plugin/notes/notes.js', async: true },
{ src: 'plugin/highlight/highlight.js', async: true, languages: ["Python"] },
{ src: 'plugin/math/math.js', async: true },
{ src: 'external-plugins/chalkboard/chalkboard.js' },
//{ src: 'external-plugins/menu/menu.js'},
{ src: 'node_modules/reveal.js-menu/menu.js' }
]
});
</script>
<script type="text/bibliography">
@article{gregor2015draw,
title={DRAW: A recurrent neural network for image generation},
author={Gregor, Karol and Danihelka, Ivo and Graves, Alex and Rezende, Danilo Jimenez and Wierstra, Daan},
journal={arXivreprint arXiv:1502.04623},
year={2015},
url={https://arxiv.org/pdf/1502.04623.pdf}
}
</script>
</body>
</html>