-
Notifications
You must be signed in to change notification settings - Fork 0
/
cs8850_23_CNN.html
1046 lines (930 loc) · 49.3 KB
/
cs8850_23_CNN.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
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/font-awesome/4.5.0/css/font-awesome.min.css">
<!-- <link href="https://stackpath.bootstrapcdn.com/font-awesome/4.7.0/css/font-awesome.min.css" rel="stylesheet"/> -->
<script src="lib/colorbrewer.v1.min.js" charset="utf-8"></script>
<script src="lib/colorStringStandalone.js" charset="utf-8"></script>
<script type="text/javascript" src="lib/jquery-2.2.4.min.js"></script>
<title>Advanced Machine Learning</title>
<meta name="description" content="CS8850 GSU class">
<meta name="author" content="Sergey M Plis">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="black-translucent">
<link rel="stylesheet" href="dist/reset.css">
<link rel="stylesheet" href="dist/reveal.css">
<!-- Code syntax highlighting -->
<link rel="stylesheet" href="plugin/highlight/monokai.css" id="highlight-theme">
<!-- <link rel="stylesheet" href="lib/css/zenburn.css"> -->
<link rel="stylesheet" href="css/custom.css">
<link rel="stylesheet" href="dist/theme/aml.css" id="theme">
<!-- 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.scss';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
</head>
<body>
<div class="reveal">
<!-- In between the <div="reveal"> and the <div class="slides">-->
<!-- <header style="position: absolute; top: 10px; left: 100px; z-index: 500; font-size:100px;background-color: rgba(0,0,0,0); text-align: center !important"></header> -->
<!-- In between the <div="reveal"> and the <div class="slides">-->
<!-- Any section element inside of this container is displayed as a slide -->
<div class="slides">
<section>
<section>
<p>
<h2>Advanced Machine Learning</h2>
<h3>24: Convolutional Neural Networks</h3>
<p>
</section>
<section>
<h3>Outline for the lecture</h3>
<ul>
<li class="fragment roll-in"> History of CNNs
<li class="fragment roll-in"> Bulding Blocks
<li class="fragment roll-in"> Skip Connections
<li class="fragment roll-in"> Fully Convolutional Neural Nets
<li class="fragment roll-in"> Semantic Segmentation with Twists
<li class="fragment roll-in"> (even more) Advanced Uses of CNN
</ul>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>Convolutions what?</h2>
</section>
<section data-background="figures/convolution.gif" data-background-size="contain">
</section>
<section data-background="figures/convolution_2.gif" data-background-size="contain" data-background-repeat="repeat">
</section>
<section data-background="figures/convolution_3.gif" data-background-size="contain" data-background-repeat="repeat">
</section>
<section data-background="figures/convolution_4.gif" data-background-size="contain" data-background-repeat="repeat">
</section>
<section data-background="figures/convolution_matching.gif" data-background-size="contain" data-background-repeat="repeat">
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>History of CNNs</h2>
</section>
<section>
<h2>Cat's brain 1962 (Hubel and Wiesel)</h2>
<img width="70%"
src="figures/cats_brain.png" alt="cat's brain">
<div class="slide-footer">
<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1359523/pdf/jphysiol01247-0121.pdf">Receptive fields, binocular interaction and functional architecture in the cat's visual cortex</a>
</div>
</section>
<section>
<div id="header-right" style="margin-right: -100px;">
<img width="130"
src="figures/fukushima.jpg" alt="fukushima">
</div>
<h2>Fukushima's Neurocognitron 1979</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="900"
src="figures/neurocognitron_.png" alt="neurocognitron">
<div class="slide-footer">
<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1359523/pdf/jphysiol01247-0121.pdf">Neocognitron - A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position</a>
</div>
</section>
<section>
<div id="header-right" style="margin-right: -100px;">
<img width="130"
src="figures/waibel.jpg" alt="waibel">
</div>
<h2>Time Delay Neural Network 1989</h2>
<img style="margin-top: -20px;" width="40%"
src="figures/tdnn_.png" alt="tdnn">
<div class="slide-footer">
<a href="https://www.cs.toronto.edu/~fritz/absps/waibelTDNN.pdf">Phoneme Recognition Using Time Delay Neural Networks</a>
</div>
<aside class="notes">
Note 1D channels!
</aside>
</section>
<section>
<div id="header-right">
<img width="130"
src="figures/lecun.png" alt="lecun">
</div>
<h2>CNN 1989</h2>
<img style="margin-top: -50px;" width="73%"
src="figures/cnn.png" alt="cnn">
<div class="slide-footer">
<a href="http://yann.lecun.org/exdb/publis/pdf/lecun-89e.pdf">Backpropagation Applied to Handwritten Zip Code Recognition</a>
</div>
<aside class="notes">
Note undersampling to reduce layers!
</aside>
</section>
<section>
<div id="header-right">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="130"
src="figures/lecun.png" alt="lecun">
</div>
<h2>CNN 1998</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/lenet.png" alt="lenet">
<div class="slide-footer">
<a href="http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf">Gradient-Based Learning applied to Document Recognition</a>
</div>
<aside class="notes">
Note subsampling! No pooling yet!
</aside>
</section>
<section>
<div id="header-right">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="130"
src="figures/dancirican.jpg" alt="cirican">
</div>
<h2>CNN+GPU+MaxPooling 2011</h2>
<img style="margin-top: -40px;" width="70%"
src="figures/gpunet.png" alt="gpunet">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1102.0183.pdf">High Performance Neural Network for Visual Object Classification</a>
</div>
<aside class="notes">
Everything is here. Max Pooling, GPU (60x speedup compared to CPU), lots of data etc.<br>
Chinese handwriting recognition 2011 winner.<br>
Traffic sign recognition winner 2011 <br>
Neuronal Membrane segmentation 2012 <br>
</aside>
</section>
<section>
<div id="header-right">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="130"
src="figures/alex.png" alt="Alex">
</div>
<h2>AlexNet 2012</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/alexnet.png" alt="alexnet">
<div class="slide-footer">
<a href="https://www.cs.toronto.edu/~fritz/absps/imagenet.pdf">ImageNet Classification with Deep ConvolutionalNeural Networks</a>
</div>
<aside class="notes">
</aside>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>CNN: bulding blocks</h2>
<div class="slide-footer">
<a href="http://deeplearning.net/software/theano_versions/dev/tutorial/conv_arithmetic.html">some images are from Theano documentation</a>
</div>
</section>
<section>
<h2>Convolving a kernel with an image</h2>
<row>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_convo1.gif" alt="convolution">
\[
\left(
\begin{array}{ccc}
0 & 1 & 2 \\
2 & 2 & 0 \\
0 & 1 & 2 \\
\end{array}
\right)
\]
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_convo2.gif" alt="convolution 2">
</col50>
</row>
</section>
<section>
<h2>Convolving a kernel with an image</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_flat_conv.svg" alt="convolution">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1603.07285.pdf">A guide to convolution arithmetic for deep learning</a>
</div>
</section>
<section>
<h2>Padding and symmetries</h2>
<row>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_same_pad.gif" alt="same pad">
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_full_pad.gif" alt="full pad">
</col50>
</row>
</section>
<section>
<h2>Padding and symmetries</h2>
<img style="margin-top: -35px;" width="85%"
src="figures/CNN_padded.svg" alt="convolution">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1603.07285.pdf">A guide to convolution arithmetic for deep learning</a>
</div>
</section>
<section>
<h2>How do the channels look?</h2>
<div style="position:relative; width:800px; height:800px; margin:0 auto;">
<img class="fragment current-visible" data-transition="slide fade-out" data-fragment-index="0" width="600" src="figures/CNN_original.png" style="position:absolute;top:0;left:0;border:0; box-shadow: 0px 0px 0px rgba(255, 255, 255, 255);" />
<img class="fragment current-visible" data-transition="slide fade-out" data-fragment-index="1" width="600" src="figures/CNN_layer1.png" style="position:absolute;top:0;left:0;border:0; box-shadow: 0px 0px 0px rgba(255, 255, 255, 255);" />
<img class="fragment current-visible" data-transition="slide fade-out" data-fragment-index="2" width="600" src="figures/CNN_layer2.png" style="position:absolute;top:0;left:0;border:0; box-shadow: 0px 0px 0px rgba(255, 255, 255, 255);" />
<img class="fragment current-visible" data-transition="slide fade-out" data-fragment-index="3" width="600" src="figures/CNN_layer3.png" style="position:absolute;top:0;left:0;border:0; box-shadow: 0px 0px 0px rgba(255, 255, 255, 255);" />
</div>
</section>
<section>
<h2>Pooling: maxpooling</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_maxpooling.gif" alt="max pooling">
</section>
<section>
<h2>Pooling: maxpooling</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_flat_maxpooling.svg" alt="max pooling">
</section>
<section>
<h2>Pooling: average</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_avpooling.svg" alt="average pooling">
</section>
<section>
<h2>How do we produce a class prediction?</h2>
</section>
<section>
<h2>One-convolution</h2>
<row>
<col50>
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_one_conv.gif" alt="one_conv">
</col50>
</row>
</section>
<section>
<div id="header-right">
<img width="300"
src="figures/CNN_upconvolution.gif" alt="gif upconv">
</div>
<h2>Upconvolution</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_upconvolution.png" alt="upconv">
<div class="slide-footer">
<a href="https://datascience.stackexchange.com/a/20176">Image from a comment on stackexchange</a>
</div>
</section>
<section>
<h2>Dilated convolution</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="60%"
src="figures/CNN_dilation.gif" alt="dilated">
</section>
<section data-background-iframe="https://www.youtube.com/embed/HnWIHWFbuUQ?autoplay=1&controls=0&rel=0&modestbranding=1&showinfo=0">
<h2 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1">Play with a simulator</h2>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1"><a href="https://www.youtube.com/watch?v=HnWIHWFbuUQ&featu" target="_blank">Video</a></h3>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1"><a href="https://poloclub.github.io/cnn-explainer/" target="_blank">Demo</a></h3>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1"><a href="https://github.com/poloclub/cnn-explainer" target="_blank">GitHub</a></h3>
<div class="slide-footer">
<a href="https://arxiv.org/abs/2004.15004v1">CNN Explainer: Learning Convolutional Neural Networks with Interactive Visualization</a>
</div>
</section>
<section>
<h2>Basic building blocks</h2>
<ol>
<li> Convolution with a filter
<li> Zero Padding
<li> Channels and channel-kernel relationship
<li> Pooling (max and average)
<li> Moving from convolution layers to predictions
<li> One convolution
<li> Upconvolution
<li> Dilated convolution
</ol>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>Skip connections</h2>
<aside class="notes">
Thanks to the reverse mode accumulation of Bert Speelpenning, (pause) after the forward pass backpropagation works backward along the paths in the computation graph. Normally, each edge denotes a computation, such as linear transformation followed by a slight nonlinearity. However, there is an advantage in excluding computation on some edges simply using an identity transform instead. A trick, that although, as everything under the sun, been known for a while, recently rediscovered to be effective in training very deep deep-learning models.<br>In this section I will walk you through the rediscovery and along the way introduce another transformative concept of gates.
</aside>
</section>
<section>
<h2>Dark knowledge</h2>
<row>
<col50>
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/shadow-of-the-vampire.jpg" alt="vampire">
</col50>
</row>
<div class="slide-footer">
<a href="https://arxiv.org/abs/1312.6184">Do Deep Nets Really Need to be Deep?</a><br>
<a href="https://arxiv.org/abs/1503.02531">Distilling the Knowledge in a Neural Network</a>
<a href="https://arxiv.org/abs/1412.6550">FitNets: Hints for Thin Deep Nets</a>
</div>
<aside class="notes">
It has been conjectured quite early that deeper models should be more powerful than shallow, but we were unable to train them effectively no matter how hard we've tried. Various pre-training approaches, when parameters of each layer are initialized in a smart way, were helping only to a point. The difficulties were discouraging and we did not know how to move forward with thin and deep networks.<br>
Until in 2013 Rich Caruana observed the following effect: (Explain the mimic nets for teacher and student training) <br>
Geoff Hinton - called the grandfather of deep learning by some - came up with explanation, which he termed: dark knowledge (explain hard labels and soft labels) <br>
and the team of Yoshia Bengio has used this observation for training deep and thin networks they called fitnets<br>
yet, to some, this two stage teacher+student process seemed a bit cumbersome.
</aside>
</section>
<section>
<h3>Highway networks (May 2015 on arxiv)</h2>
<ul style="list-style-type: none; font-size: 32px;">
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H)
$$
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H) \odot T(\vec{x}, \bm{W}_T) + \vec{x} \odot C(\vec{x}, \bm{W}_C)
$$
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H) \odot T(\vec{x}, \bm{W}_T) + \vec{x} \odot (1 - T(\vec{x}, \bm{W}_T))
$$
<li class="fragment roll-in">
$$
\vec{y} =
\left\{
\begin{array}{ll}
\vec{x} & \mbox{if }\;\;T(\vec{x}, \bm{W}_T)=0,\\
H(\vec{x}, \bm{W}_H) & \mbox{if }\;\;T(\vec{x}, \bm{W}_T)=1
\end{array}
\right.
$$
<li class="fragment roll-in"> What if untrained gate is always open and does not let gradients flow?
<li class="fragment roll-in"> <b>Initialize gate biases to large negative values!</b>
</ul>
<div class="slide-footer" style="text-align: left;">
<a href="https://arxiv.org/abs/1505.00387">Highway Networks</a><br>
<a href="https://arxiv.org/abs/1507.06228">Training Very Deep Networks</a>
</div>
<aside class="notes">
In 2015 Rupesh Srivastawa, Klaus Greff and Yurgen Schmidhuber made the following, a bit elaborate, observation. If we embellish the regular layer in a feed forward neural network with a transfer and carry gates, we will enable gradients flow along the entire depth of a deep network.<br>
Let's see how they have done it.
</aside>
</section>
<section data-fullscreen>
<alert>Train models with 100 of layers instead of just 10 before</alert>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_HW_performance.png" alt="HWN">
<aside class="notes">
The networks that they have trained using traditional layers and highway layers were of depth 10, 20, 50, and a 100. Plain networks would train increasingly poorly with depth, while highway networks retained ability to train. Importantly the results were comparable with the ones of fitnets (trained by mimicing a teacher).
</aside>
</section>
<section>
<h2>Residual Networks (block)</h2>
<row>
<col50>
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_block.svg" alt="resnet block">
</col50>
</row>
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a>
</div>
</section>
<section>
<h2>Residual Networks (full)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_whole.svg" alt="resnet whole">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a>
</div>
</section>
<section>
<h2>Residual Networks (performance)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_perf.svg" alt="resnet perf">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a>
</div>
</section>
<section>
<h2>Error surface effect of skip connection</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/skip_smooth.png" alt="landscape">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a><br>
<a href="https://arxiv.org/pdf/1505.00387.pdf">Highway Networks</a><br>
<a href="https://papers.nips.cc/paper/7875-visualizing-the-loss-landscape-of-neural-nets.pdf">Visualizing the Loss Landscape of Neural Nets</a>
</div>
<aside class="notes">
</aside>
</section>
<section>
<h2>Dense Networks (architecture)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="80%"
src="figures/CNN_densenet.svg" alt="dense diagram">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1608.06993">Densely Connected Convolutional Networks</a>
</div>
<aside class="notes">
</aside>
</section>
<section>
<h2>Dense Networks (effect)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/densenet.png" alt="dense">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1608.06993">Densely Connected Convolutional Networks</a>
</div>
<aside class="notes">
</aside>
</section>
<section>
<h2>Take Away Concepts</h2>
<ol>
<li> Skip connections
<li> Gates
</ol>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>Fully convolutional networks</h2>
</section>
<section data-background-iframe="https://www.youtube.com/embed/xr_2dwZDH6U?autoplay=1&controls=0&rel=0&modestbranding=1&showinfo=0">
<h2 style="color:#839496;">The task of Semantic segmentation</h2>
</section>
<section data-fullscreen>
<h2>Semantic segmentation task</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_semantic_bikes.png" alt="fikes">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1411.4038">Fully Convolutional Networks for Semantic Segmentation</a><br>
</div>
</section>
<section>
<h3>Replacing feed forward with convolutional</h3>
<ul style="list-style-type: none; margin: 0px 0;">
<li class="fragment roll-in">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="800"
src="figures/CNN_semantic_cc_1.png" alt="cc 1">
<li class="fragment roll-in">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="800"
src="figures/CNN_semantic_cc_2.png" alt="cc 1">
<li class="fragment roll-in">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="800"
src="figures/CNN_semantic_cc_3.png" alt="cc 1">
</ul>
</section>
<section data-fullscreen>
<h2>Fully Convolutional Model (2014)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_semantic_ff.png" alt="final model">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1411.4038" target="_blank">Fully Convolutional Networks for Semantic Segmentation</a><br>
</div>
</section>
<section data-fullscreen>
<h2>Examples</h2>
<img width="50%"
src="figures/CNN_semantic_examples.png" alt="final model">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1411.4038" target="_blank">Fully Convolutional Networks for Semantic Segmentation</a><br>
</div>
</section>
<section>
<h2>Take Away Point</h2>
<ol>
<li> When target and input have the same dimension it may be better to use convolution everywhere.
</ol>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>Semantic segmentation with twists</h2>
</section>
<section>
<h2>deep learning standard: U-net</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="250"
src="figures/unet_arc.png" alt="david">
<p>
<div class='slide-footer'>
Ronneberger et al. U-net: Convolutional networks
for biomedical image segmentation. MICCAI 2015<br>
Çiçek et al. 3D U-Net: learning dense volumetric
segmentation from sparse annotation. MICCAI 2016
</div>
</section>
<section>
<h2>deep learning standard: U-net</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="250"
src="figures/unet_table.png" alt="david">
<p>
<div class='slide-footer'>
Ronneberger et al. U-net: Convolutional networks for
biomedical image segmentation. MICCAI 2015<br>
Çiçek et al. 3D U-Net: learning dense volumetric
segmentation from sparse annotation. MICCAI 2016
</div>
</section>
<section>
<h2>comparison on the brain segmentation task</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="250"
src="figures/tasks.png" alt="david">
</section>
<section>
<h2>state of the art: freesurfer</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="250"
src="figures/CNN_freesurfer.png" alt="david">
<br>
<div class='slide-footer'>
Dale et al. Cortical surface-based analysis. I. Segmentation and
surface reconstruction. Neuroimage 1999
</div>
</section>
<section>
<h2>Meshnet</h2>
<img src="figures/dilation.png" style="border:0; box-shadow: 0px
0px 0px rgba(150, 150, 255, 0.8);float: left;"
alt="Websocket" width="40%">
<img src="figures/mn1.png" style="border:0; box-shadow: 0px
0px 0px rgba(150, 150, 255, 0.8);float: right;"
alt="Websocket" width="35%"
>
<div class="ulist">
<ul style="width: 60%; float: left;">
<li>Gray and White matter</li>
<li>FreeSurfer for ground truth</li>
<li>T1 MRIs from HCP</li>
<li>GitHub <br>
<small>
<a href="https://github.com/Entodi/MeshNet">
https://github.com/Entodi/MeshNet</a>
</small>
</li>
</ul>
</div>
<br>
<div class="slide-footer">
Fedorov et al. End-to-end learning of brain tissue segmentation
from imperfect labeling. IJCNN 2017
</div>
</section>
<section>
<h2>Meshnet</h2>
<img src="figures/mntable.png" style="border:0; box-shadow: 0px
0px 0px rgba(150, 150, 255, 0.8);float: right;"
alt="Websocket" width="60%"
>
<div class="ulist">
<ul style="width: 35%; float: left;">
<li>72516 vs. 23523355</li>
<li>600kb vs. 2Gb</li>
</ul>
</div>
<p>
<br>
<div class="slide-footer">
Fedorov et al. End-to-end learning of brain tissue segmentation
from imperfect labeling. IJCNN 2017
</div>
</section>
<section>
<h2>Meshnet</h2>
<img src="figures/mnpipeline.png" style="border:0; box-shadow: 0px
0px 0px rgba(150, 150, 255, 0.8);"
alt="Websocket" class="stretch"
>
<br>
<div class="slide-footer">
Fedorov et al. End-to-end learning of brain tissue segmentation
from imperfect labeling. IJCNN 2017
</div>
</section>
<section>
<h2>Meshnet</h2>
<img src="figures/mnexample.png" style="border:0; box-shadow: 0px
0px 0px rgba(150, 150, 255, 0.8);"
alt="Websocket" class="stretch"
>
<br>
<div class="slide-footer">
Fedorov et al. End-to-end learning of brain tissue segmentation
from imperfect labeling. IJCNN 2017
</div>
</section>
<section>
<h3>(often) better than the teacher</h3>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="700"
src="figures/MN_examples.png" alt="loop">
</section>
<section>
<h3>multimodal is straightforward</h3>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="700"
src="figures/MeshNet_multi.png" alt="loop">
</section>
<section>
<h3>better than the human (sometimes)</h3>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="700"
src="figures/mnhuman.png" alt="loop">
</section>
<section>
<h3>better than U-net</h3>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);"
src="figures/mnvsunet.png" alt="loop">
</section>
<section data-background-video="figures/combined.mp4" data-background-size="contain" data-background-video-loop=true>
</section>
</section>
<!-- --------------------------------------------------------------------------->
<section>
<section>
<h2>(even more) "Advanced" uses of CNN</h2>
</section>
<section>
<h2>Masked Convolutions</h2>
<div class="slide-footer">
<a href="https://arxiv.org/abs/1601.06759">
Pixel Recurrent Neural Networks
</a><br>
<a href="https://arxiv.org/abs/1606.05328">
Conditional Image Generation with PixelCNN Decoders
</a>
</div>
</section>
<section>
<h2>PixelCNNs</h2>
<div class="slide-footer">
<a href="https://arxiv.org/abs/1601.06759">
Pixel Recurrent Neural Networks
</a><br>
<a href="https://arxiv.org/abs/1606.05328">
Conditional Image Generation with PixelCNN Decoders
</a>
</div>
</section>
<section>
<h2>Wavenet: $\ge$16kHz audio</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);"
src="figures/CNN_audio_zoom.gif" alt="loop">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1609.03499">
WaveNet: A Generative Model for Raw Audio
</a>
</div>
</section>
<section>
<h2>Wavenet: sample by sample</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);"
src="figures/CNN_wavenet.gif" alt="loop">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1609.03499">
WaveNet: A Generative Model for Raw Audio
</a>
</div>
</section>
<section>
<h2>Wavenet: conditioned on text</h2>
<row>
<col50>
<table style="font-size:28px">
<tr>
<th>Model</th>
<th>"The blue lagoon..."</th>
</tr>
<tr>
<td>Parametric</td>
<td><audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/parametric-1.wav"></audio></td>
</tr>
<tr>
<td>Concatenative</td>
<td>
<audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/concatenative-1.wav"></audio>
</td>
</tr>
<tr>
<td>Wavenet</td>
<td>
<audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/wavenet-1.wav"></audio>
</td>
</tr>
</table>
</col50>
<col50>
<table style="font-size:28px">
<tr>
<th>Model</th>
<th>"English poetry and ..."</th>
</tr>
<tr>
<td>Parametric</td>
<td><audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/parametric-2.wav"></audio></td>
</tr>
<tr>
<td>Concatenative</td>
<td>
<audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/concatenative-2.wav"></audio>
</td>
</tr>
<tr>
<td>Wavenet</td>
<td>
<audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/US%20English/wavenet-2.wav"></audio>
</td>
</tr>
</table>
</col50>
</row>
<audio controls="" src="https://storage.googleapis.com/deepmind-media/research/WaveNet/Speaker%201/speaker-2.wav"></audio>
<div class="slide-footer">
<a href="https://arxiv.org/abs/1609.03499">
WaveNet: A Generative Model for Raw Audio
</a>
</div>
</section>
<section data-fullscreen>
<h2>Deformable Convolutions</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);"
src="figures/CNN_deformable.svg" alt="defrmable">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1703.06211">
Deformable Convolutions
</a>
</div>
</section>
<section data-fullscreen>
<h2>Deformable Convolutions</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);"
src="figures/CNN_deformable_conv_demo1.png" alt="defrmable">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1703.06211">
Deformable Convolutions
</a>
</div>
</section>
<section>
<h2>Take Away Points</h2>
<ol>
<li> Masked convolution
<li> Pixel based generation
<li> Deformable convolution (can be rotation invariant)
</ol>
</section>
</section>
</div>
</div>
<script src="dist/reveal.js"></script>
<link rel="stylesheet" href="plugin/highlight/monokai.css">
<script src="plugin/highlight/highlight.js"></script>
<script src="plugin/math/math.js"></script>
<script src="plugin/chalkboard/plugin.js"></script>
<script src="plugin/notes/notes.js"></script>
<script src="plugin/zoom/zoom.js"></script>
<script src="plugin/fullscreen/fullscreen.js"></script>
<script src="plugin/menu/menu.js"></script>
<script src="plugin/verticator/verticator.js"></script>
<link rel="stylesheet" href="plugin/verticator/verticator.css">
<script>
// Full list of configuration options available at:
// https://github.com/hakimel/reveal.js#configuration
Reveal.initialize({
// history: true,
width: 960,
height: 700,
center: true,
hash: true,
controls: false,
keyboard: true,
margin: 0.05,
overview: true,
transition: 'slide', // Transition style: none/fade/slide/convex/concave/zoom
transitionSpeed: 'slow', // Transition speed: default/fast/slow
menu: {
themes: false,
openSlideNumber: true,
openButton: false,
},
chalkboard: {
boardmarkerWidth: 1,
chalkWidth: 2,
chalkEffect: 1,
toggleNotesButton: false,
toggleChalkboardButton: false,
slideWidth: Reveal.width,
slideHeight: Reveal.height,
// src: "chalkboards/chalkboard_em2.json",
readOnly: false,
theme: "blackboard",
eraser: { src: "plugin/chalkboard/img/sponge.png", radius: 30},
},
math: {
mathjax: 'https://cdn.jsdelivr.net/gh/mathjax/[email protected]/MathJax.js',
config: 'TeX-AMS_SVG-full',
// pass other options into `MathJax.Hub.Config()`
TeX: {
Macros: {
RR: '\\mathbb{R}',
PP: '\\mathbb{P}',
EE: '\\mathbb{E}',
NN: '\\mathbb{N}',
vth: '\\vec{\\theta}',
loss: '{\\cal l}',
hclass: '{\\cal H}',
CD: '{\\cal D}',
def: '\\stackrel{\\text{def}}{=}',
pag: ['\\text{pa}_{{\cal G}^{#1}}(#2)}', 2],
vec: ['\\boldsymbol{\\mathbf #1}', 1],
set: [ '\\left\\{#1 \\; : \\; #2\\right\\}', 2 ],
bm: ['\\boldsymbol{\\mathbf #1}', 1],
argmin: ['\\operatorname\{arg\\,min\\,\}'],
argmax: ['\\operatorname\{arg\\,max\\,\}'],
prob: ["\\mbox{#1$\\left(#2\\right)$}", 2],
},
loader: {load: ['[tex]/color']},
extensions: ["color.js"],
tex: {packages: {'[+]': ['color']}},
svg: {
fontCache: 'global'
}
}
},
plugins: [ Verticator, RevealMath, RevealChalkboard, RevealHighlight, RevealNotes, RevealZoom, RevealMenu ],
});
Reveal.configure({ fragments: true }); // set false when developing to see everything at once
Reveal.configure({ slideNumber: true });
//Reveal.configure({ history: true });
Reveal.configure({ slideNumber: 'c / t' });
Reveal.addEventListener( 'darkside', function() {
document.getElementById('theme').setAttribute('href','dist/theme/aml_dark.css');
}, false );
Reveal.addEventListener( 'brightside', function() {
document.getElementById('theme').setAttribute('href','dist/theme/aml.css');
}, false );
</script>
<style type="text/css">