-
Notifications
You must be signed in to change notification settings - Fork 4
/
index.html
618 lines (533 loc) · 22.9 KB
/
index.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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<title>NeuralNILM</title>
<meta name="description" content="Deep Neural Networks applied to Energy Disaggregation">
<meta name="author" content="Jack Kelly">
<meta name="apple-mobile-web-app-capable" content="yes" />
<meta name="apple-mobile-web-app-status-bar-style" content="black-translucent" />
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no, minimal-ui">
<link rel="stylesheet" href="bower_components/reveal.js/css/reveal.css">
<link rel="stylesheet" href="bower_components/reveal.js/css/theme/black.css" id="theme">
<link rel="stylesheet" href="css/jack.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 ) ? 'bower_components/reveal.js/css/print/pdf.css' : 'bower_components/reveal.js/css/print/paper.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
<!--[if lt IE 9]>
<script src="bower_components/reveal.js/lib/js/html5shiv.js"></script>
<![endif]-->
</head>
<body>
<div class="reveal">
<!-- Any section element inside of this container is displayed as a slide -->
<div class="slides">
<section>
<h1>Neural NILM</h1>
<h3>Deep Neural Networks Applied To<br>Energy
Disaggregation</h3>
<p>
<small>
<a href="http://jack-kelly.com">Jack Kelly</a>
& <a href="http://www.doc.ic.ac.uk/~wjk/">William
Knottenbelt</a>
<br/>
Imperial College London<br/>
(<em>Swipe or press right-arrow on your keyboard to change slides</em>)
</small></p>
<p>
<small>
</small>
</p>
</section>
<section>
<h2>Energy Disaggregation</h2>
<img src="images/disag.png" height=550>
</section>
<section>
<h2>Aim: Itemised Energy Bills</h2>
<img src="images/itemised_energy.png" height=550>
</section>
<section id="outline">
<h1>Outline</h1>
<ol style="width:70%">
<li class="fragment">Why use deep neural nets (DNNs) for
NILM?</li>
<!-- Describe *what* DNNs are good at. -->
<li class="fragment">How DNNs work</li>
<li class="fragment">Three DNN architectures for NILM</li>
<li class="fragment">Data augmentation</li>
<li class="fragment">Results</li>
<li class="fragment">Summary</li>
</ol>
</section>
<!-- ******************************************************
In this section, describe *what* DNNs are good at (not how).
******************************************************** -->
<section id="why-dnns-for-nilm">
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li>Why use deep neural nets (DNNs) for NILM?</li>
<li class="dim">How DNNs work</li>
<li class="dim">Three DNN architectures for NILM</li>
<li class="dim">Data augmentation</li>
<li class="dim">Results</li>
<li class="dim">Summary</li>
</ol>
</section>
<section id="washer">
<h1>Name the Appliance?</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
<span class="fragment"></span>
</section>
<section>
<img src="images/washing-machine-transparent.png" height=560
style="background-color: rgb(34,34,34); box-shadow: none;">
</section>
<section>
<h1>Face Recognition</h2>
<h2 class="fragment">Manual Feature Extraction</h3>
</section>
<section>
<div style="width: 673px; height: 598px;
border: solid white;
margin: auto;">
<img src="images/hart1.jpg"
style="margin: 0;
position: absolute;
transform: translate(-50%, 0)">
<object data="images/hart_circles.svg"
class="fragment"
style="margin: 0;
position: absolute;
transform: translate(-50%,0);"
type="image/svg+xml"></object>
</div>
<p class="tiny fragment">
<a href="http://www.georgehart.com/research/hartbiog.html">
georgehart.com/research/hartbiog.html</a>
</p>
</section>
<section>
<img src="images/hart2.jpg" height=600>
<p class="tiny">
<a href="http://scgp.stonybrook.edu/archives/8516">
scgp.stonybrook.edu/archives/8516</a>
</p>
</section>
<!--
<section>
<h1>SIFT</h1>
<p>Scale-Invariant Feature Transform</p>
<img src="images/sift_keypoints.jpg">
<p class="tiny">
Image from: <a href="http://opencv-python-tutroals.readthedocs.org/en/latest/py_tutorials/py_feature2d/py_sift_intro/py_sift_intro.html">
OpenCV-Python Tutorials: Introduction to SIFT</a>
</p>
<p class="tiny">
David G. Lowe. <a href="http://dx.doi.org/10.1109/ICCV.1999.790410">Object recognition from local
scale-invariant
features</a>. ICCV (1999)
<br>Cited by 10,106 papers!
</p>
</section>
-->
<section>
<h1>Deep Neural Nets</h1>
<h2>Automatic Feature Learning</h2>
</section>
<section>
<img src="images/faces.png">
<img class="fragment" src="images/cars.png">
<img class="fragment" src="images/chairs.png">
</section>
<section>
<p><strong>ImageNet</strong> Large Scale Visual Recognition
Challenge (ILSVRC)</p>
<img src="images/imagenet.png" height=550>
<p class="tiny">From: Krizhevsky, Sutskever &
Hinton. <a href="http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks">ImageNet Classification with Deep
Convolutional Neural Networks</a>. NIPS (2012)
</p>
</section>
<section>
<img src="images/imagenetResults.png" width="90%">
<p class="tiny">Image from <a href="http://devblogs.nvidia.com/parallelforall/mocha-jl-deep-learning-julia/">devblogs.nvidia.com</a></p>
</section>
<section>
<p><strong>Krizhevsky <em>et al.</em>'s DNN Results on ImageNet 2012</strong></p>
<img src="images/krizhevskyResults.png" height=550>
<p class="tiny">Krizhevsky, Sutskever &
Hinton. <a href="http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks">ImageNet Classification with Deep
Convolutional Neural Networks</a>. NIPS (2012)
</p>
</section>
<!-- ******************************************************
*HOW* DNNS work...
******************************************************** -->
<section id="how-dnns-work">
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li class="dim">Why use deep neural nets (DNNs) for NILM?</li>
<li>How DNNs work</li>
<li class="dim">Three DNN architectures for NILM</li>
<li class="dim">Data augmentation</li>
<li class="dim">Results</li>
<li class="dim">Summary</li>
</ol>
</section>
<section>
<h1>The Artificial Neuron</h1>
<img src="images/ArtificialNeuronModel.png">
<p class="tiny">Image adapted from <a href="https://en.wikibooks.org/wiki/File:ArtificialNeuronModel_english.png">WikiMedia Commons image by Chrislb</a></p>
<!-- TODO:
** D3 (low priority): flow of info through neuron **
-->
</section>
<section id="neuralnet">
<h1>Feed Forward Nets</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
</section>
<section>
<p><strong>Krizhevsky <em>et al.</em>'s Architecture for ImageNet 2012</strong></p>
<img src="images/krizhevskyArchitecture.png">
<p class="tiny">Krizhevsky, Sutskever &
Hinton. <a href="http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks">ImageNet Classification with Deep
Convolutional Neural Networks</a>. NIPS (2012)
</p>
</section>
<section>
<h1>Training</h1>
<object data="images/training.svg" type="image/svg+xml"></object>
<!--
Need to learn the weights and biases.
With 100M learnable parameters, it's not tractable to do an
exhaustive search or a random search.
So use gradient descent.
Back prop.
** D3 (low priority): gradient descent & back-prop **
-->
</section>
<section id="autoencoder">
<h1>Autoencoders</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
</section>
<section>
<h2>Autoencoder Examples</h2>
<div style="height: 600px;">
<img class="fragment current-visible"
style="position:absolute; transform: translate(-50%, 0);"
src="images/HintonAE.png" height=560>
<img class="fragment current-visible"
style="position:absolute; transform: translate(-50%, 50%);"
src="images/HintonAEexamples.png" width=1100>
</div>
<p class="tiny">Hinton & Salakhutdinov. <strong>Reducing the
dimensionality of data with neural
networks.</strong> <em>Science</em> (2006)</p>
</section>
<!--
<section>
<h2>Autoencoder Examples</h2>
<img src="images/learned_features.gif" height=550>
<p class="tiny">Animation by <a href="http://www.cs.toronto.edu/~ranzato/research/projects.html#sparse_coding">Marc'Aurelio Ranzato</a></p>
</section>
-->
<section>
<h1>Denoising Autoencoders</h1>
<img src="images/peppers.png">
<p class="tiny">Image from <a href="http://www.cs.toronto.edu/~ranzato/research/projects.html">Marc'Aurelio Ranzato</a></p>
<p class="tiny">Vincent <em>et al.</em> <strong>Extracting and composing robust features
with denoising autoencoders.</strong> <em>ICML</em> (2008)</p>
</section>
<section id="recurrent">
<h1>Recurrent Neural Nets</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
<span class="fragment"></span>
</section>
<section>
<h1>Recurrent Neural Nets</h1>
<p>Long Short-Term Memory (LSTM) Cells</p>
<img src="images/LSTMcell.png" height=400>
<p class="tiny">Image
from <a href="http://blog.otoro.net/page/3/">blog.otoro.net</a></p>
<p class="tiny">
Hochreiter & Schmidhuber. <strong>Long short-term memory</strong>. <em>Neural Computation</em> (1997)
</p>
<!--
*** D3 (low priority): LSTM animation ***
-->
</section>
<section>
<h2>Recurrent Neural Nets</h2>
<p> Playing Volleyball :)</p>
<video height=500 controls>
<source src="images/volleyball.mp4" type="video/mp4">
</video>
<p class="tiny">
By <a href="http://blog.otoro.net/2015/03/28/neural-slime-volleyball/">
hardmaru / ōtoro / 大トロ
</a>
</p>
</section>
<!-- ******************************************************
Three DNN ARCHITECTURES FOR NILM
******************************************************** -->
<section id="dnn-architectures-for-nilm">
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li class="dim">Why use deep neural nets (DNNs) for NILM?</li>
<li class="dim">How DNNs work</li>
<li>Three DNN architectures for NILM
<ol>
<li>Recurrent Neural Nets (LSTM)</li>
<li>Denoising Autoencoder</li>
<li>'Bounding rectangle' around the target</li>
</ol>
</li>
<li class="dim">Data augmentation</li>
<li class="dim">Results</li>
<li class="dim">Summary</li>
</ol>
</section>
<section id="rnns-for-nilm">
<h1>Recurrent Neural Nets</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
<span class="fragment"></span>
<span class="fragment"></span>
<span class="fragment"></span>
<!--
***** D3 (low priority) ******
Highlight the connections as data flows through, especially
RNN connections.
-->
</section>
<section id="autoencoder-for-nilm">
<h1>Denoising Autoencoders</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
<span class="fragment"></span>
<span class="fragment"></span>
<!--
***** D3 (low priority) ******
Highlight the connections as data flows through.
-->
</section>
<section id="rectangles-for-nilm">
<h1>Bounding Rectangle</h1>
<p class="placeholder"></p>
<span class="fragment"></span>
<span class="fragment"></span>
<span class="fragment"></span>
<!--
***** D3 (low priority) ******
Highlight the connections as data flows through.
-->
</section>
<!-- ******************************************************
DATA AUGMENTATION
******************************************************** -->
<section>
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li class="dim">Why use deep neural nets (DNNs) for NILM?</li>
<li class="dim">How DNNs work</li>
<li class="dim">Three DNN architectures for NILM</li>
<li>Data augmentation</li>
<li class="dim">Results</li>
<li class="dim">Summary</li>
</ol>
</section>
<section>
<h1>DNNs need lots of data!</h1>
</section>
<section>
<p><strong>Data Augmentation for Images of Plakton</strong></p>
<div style="height: 600px;">
<div class="fragment current-visible" style="position:absolute;">
<p style="position:absolute; transform: translate(100px, 200px);">
Raw
</p>
<img src="images/plankton_noaug.png"
style="transform: translate(60%, 0)"
height=550>
</div>
<div class="fragment current-visible" style="position:absolute;" >
<p style="position:absolute; transform: translate(50px, 200px);">
Augmented
</p>
<img src="images/plankton_augmented.png"
style="transform: translate(60%, 0)"
height=550>
</div>
<div style="position:absolute; transform: translate(70px, 590px);">
<p class="tiny" style="text-align:center;">
From <a href="http://benanne.github.io/2015/03/17/plankton.html">≋ Deep Sea ≋ team (Dieleman <em>et al.</em>)</a> on
<a href="https://www.kaggle.com/c/datasciencebowl">
Kaggle National Data Science Bowl Plankton competition
</a>
</p>
</div>
</div>
</section>
<section>
<h3>Data Augmentation for NILM</h3>
<ul>
<li class="fragment">Extract individual appliance
activations from real data</li>
<li class="fragment">For each generated example:
<ul>
<li class="fragment">Randomly pick which appliances
to include</li>
<li class="fragment">Randomly pick individual
activations</li>
<li class="fragment">Randomly align activations</li>
</ul>
</li>
</ul>
<!-- TODO:
*** D3 (medium priority): illustrate data augmentation ***
Maybe just a flow chart: start with real appliance-level
data, extract activations into a set per appliance type,
etc...
Show several examples?
-->
</section>
<!-- ******************************************************
RESULTS
******************************************************** -->
<section>
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li class="dim">Why use deep neural nets (DNNs) for NILM?</li>
<li class="dim">How DNNs work</li>
<li class="dim">Three DNN architectures for NILM</li>
<li class="dim">Data augmentation</li>
<li>Results</li>
<li class="dim">Summary</li>
</ol>
</section>
<section>
<h3>Example Output</h3>
<div style="height: 600px; width: 1000px;">
<div class="fragment current-visible" style="position:absolute;">
<p>LSTM</p>
<object data="images/net_output_LSTM.svg"
type="image/svg+xml" width=1000></object>
</div>
<div class="fragment current-visible" style="position:absolute;">
<p>Autoencoder</p>
<object data="images/net_output_AE.svg"
type="image/svg+xml" width=1000></object>
</div>
<div class="fragment current-visible" style="position:absolute;">
<p>Rectangles</p>
<object data="images/net_output_rectangles.svg"
type="image/svg+xml" width=1000></object>
</div>
</div>
</section>
<section>
<object data="images/net_output_no_overlap.svg"
type="image/svg+xml" width=1000></object>
</section>
<section>
<h1>Metrics</h1>
</section>
<section>
<div style="height: 600px; width: 1000px;">
<div class="fragment current-visible"
style="position:absolute;">
<p class="tiny"
style="position:absolute; transform: translate(0, 0);">
Metrics on Seen Appliances
</p>
<object data="images/train_houses.svg"
type="image/svg+xml"
style="transform: translate(50%, 0);"
width=600></object>
</div>
<div class="fragment current-visible"
style="position:absolute;">
<p class="tiny"
style="position:absolute; transform: translate(0, 0);">
Metrics on Unseen Appliances
</p>
<object data="images/unseen_houses.svg"
type="image/svg+xml"
style="transform: translate(50%, 0);"
width=600></object>
</div>
</div>
</section>
<!-- ******************************************************
CONCLUSIONS
******************************************************** -->
<section>
<h1 class="dim">Outline</h1>
<ol style="width:70%">
<li class="dim">Why use deep neural nets (DNNs) for NILM?</li>
<li class="dim">How DNNs work</li>
<li class="dim">Three DNN architectures for NILM</li>
<li class="dim">Data augmentation</li>
<li class="dim">Results</li>
<li>Summary</li>
</ol>
</section>
<section>
<h1>Summary</h1>
<ol style="width:70%">
<li class="fragment">Developed 3 deep neural nets for NILM</li>
<li class="fragment">They perform better than NILMTK's CO
or FHMM algorithms (on UK-DALE)</li>
<li class="fragment">Code and Data available:<ol>
<li>www.doc.ic.ac.uk/~dk3810/neuralnilm</li>
<li>github.com/JackKelly/neuralnilm</li>
</ol>
</li>
<li class="fragment">Just scratched the surface!</li>
</ol>
</section>
</div>
</div>
<!-- REVEAL -->
<script src="bower_components/reveal.js/lib/js/head.min.js"></script>
<script src="bower_components/reveal.js/js/reveal.js"></script>
<!-- D3 -->
<script src="bower_components/d3/d3.js"></script>
<!-- REVEAL TO D3 -->
<script src="js/reveal-to-d3-config.js"></script>
<script src="js/reveal-to-d3.js"></script>
<script src="js/d3-common.js"></script>
<script src="js/d3-plot-power-data.js"></script>
<script src="js/d3-plot-neural-nets.js"></script>
<!-- REVEAL INIT -->
<script src="js/reveal-init.js"></script>
<!-- Google Analytics -->
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','https://www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-29040208-7', 'auto');
ga('send', 'pageview');
// From https://github.com/hakimel/reveal.js/issues/211#issuecomment-60726240
Reveal.addEventListener("slidechanged",function(event){
ga('send', 'pageview', {
'page': event.indexh,
'title': document.title + ' : Slide ' + event.indexh
});
});
</script>
</body>
</html>