-
Notifications
You must be signed in to change notification settings - Fork 3
/
index.html
568 lines (511 loc) · 15.1 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8"/>
<title></title>
<meta name="author" content="(Stefan Otte)"/>
<style type="text/css">
.underline { text-decoration: underline; }
</style>
<link rel="stylesheet" href="reveal.js/css/reveal.css"/>
<link rel="stylesheet" href="reveal.js/css/theme/serif.css" id="theme"/>
<!-- If the query includes 'print-pdf', include the PDF print sheet -->
<script>
if( window.location.search.match( /print-pdf/gi ) ) {
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = 'reveal.js/css/print/pdf.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
}
</script>
<script type="text/javascript" src="https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
</head>
<body>
<div class="reveal">
<div class="slides">
<section>
<section id="slide-orgb9877a6">
<h2 id="orgb9877a6">On Bandits and Swipes – Gamification of Search</h2>
<p>
<i>Stefan Otte</i>
</p>
<p>
<a href="https://goo.gl/8JgirR">https://goo.gl/8JgirR</a>
</p>
<div class="figure">
<p><img src="./img/qr.png" alt="qr.png" />
</p>
</div>
</section>
</section>
<section>
<section id="slide-orgaa4b314">
<h2 id="orgaa4b314">Active Learning or: How I Learned to Stop Worrying and Love Small Data</h2>
<aside class="notes">
<ul>
<li>spiritual title</li>
<li>show of hands</li>
</ul>
</aside>
</section>
</section>
<section>
<section id="slide-org5cd69f8">
<h2 id="org5cd69f8"></h2>
<p>
<i>Stefan Otte</i>
</p>
<p>
<a href="https://github.com/sotte">https://github.com/sotte</a>
</p>
</section>
<section id="slide-orga9a1740" data-background="img/pr2.jpg" data-background-size="80%">
<h3 id="orga9a1740"></h3>
</section>
<section id="slide-org4f1d273" data-background="img/um.png" data-background-size="400px">
<h3 id="org4f1d273"></h3>
</section>
</section>
<section>
<section id="slide-org2ec5f57" data-background="img/cars.png" data-background-size="100%">
<h2 id="org2ec5f57"></h2>
</section>
</section>
<section>
<section id="slide-org7cf08ea" data-background="img/cars2.png" data-background-size="100%">
<h2 id="org7cf08ea"></h2>
</section>
</section>
<section>
<section id="slide-org819e512" data-background="img/cars3.png" data-background-size="100%">
<h2 id="org819e512"></h2>
</section>
</section>
<section>
<section id="slide-org5dbfda5" data-background="img/cars4.png" data-background-size="100%">
<h2 id="org5dbfda5"></h2>
</section>
</section>
<section>
<section id="slide-org4e55206" data-background="img/cars5.png" data-background-size="100%">
<h2 id="org4e55206"></h2>
</section>
</section>
<section>
<section id="slide-orga83b672">
<h2 id="orga83b672"></h2>
<div class="figure">
<p><img src="./img/cinder.png" alt="cinder.png" />
</p>
</div>
</section>
</section>
<section>
<section id="slide-orgf6ce3be">
<h2 id="orgf6ce3be"><i>"Nothing Is Quite So Practical as a Good Theory"</i></h2>
<p>
– Kurt Lewin
</p>
</section>
</section>
<section>
<section id="slide-orgaa69438">
<h2 id="orgaa69438">Active Learning</h2>
<div class="outline-text-2" id="text-orgaa69438">
</div></section>
<section id="slide-orge47c816">
<h3 id="orge47c816"></h3>
<p>
<i>"The key idea behind <b>active learning</b> is that a machine learning algorithm can achieve <b>greater accuracy</b> with <b>fewer training labels</b> if it is allowed to <b>choose the data</b> from which it learns.</i>
</p>
</section>
<section id="slide-orgd3cfaef">
<h3 id="orgd3cfaef"></h3>
<p>
<i>An active learner may <b>pose queries</b>, usually in the form of unlabeled data instances to be labeled by an oracle (e.g., a human annotator).</i>
</p>
</section>
<section id="slide-org086cc33">
<h3 id="org086cc33"></h3>
<p>
<i>Active learning is well-motivated in many modern machine learning problems, where unlabeled data may be abundant or easily obtained, but <b>labels are difficult, time-consuming, or expensive to obtain</b>."</i>
</p>
<p>
– Burr Settles, Active Learning Literature Survey
</p>
<aside class="notes">
<ul>
<li>greater accuracy</li>
<li>fewer training labels</li>
<li>pose queries</li>
<li>when labels are expensive</li>
</ul>
</aside>
</section>
<section id="slide-org2448c1f" data-background="img/al.png" data-background-size="50%" data-background-transition="slide">
<h3 id="org2448c1f"></h3>
</section>
<section id="slide-org2c8c382" data-background="img/burr_settles.jpg" data-background-transition="slide">
<h3 id="org2c8c382"></h3>
</section>
<section id="slide-orga6a627d">
<h3 id="orga6a627d"></h3>
<p>
greater accuracy with fewer training labels
</p>
<p class="fragment (roll-in)">
<b>→ "good data<sup>TM</sup>"</b>
</p>
<p>
actively query for data
</p>
<p class="fragment (roll-in)">
<b>→ sequential decision making</b>
</p>
<aside class="notes">
<p>
the essence of active learning
</p>
</aside>
</section>
</section>
<section>
<section id="slide-org761c65c">
<h2 id="org761c65c">\({\huge \textbf{X} \rightarrow} \begin{bmatrix} cat\\ dog\\ \vdots\\ cat \end{bmatrix}\)</h2>
</section>
</section>
<section>
<section id="slide-org9a8895d">
<h2 id="org9a8895d">\({\huge \textbf{X} \rightarrow} \begin{bmatrix} ?\\ ?\\ \vdots\\ ? \end{bmatrix}\)</h2>
</section>
</section>
<section>
<section id="slide-org376e71c">
<h2 id="org376e71c">What is <b>Interesting</b>?</h2>
</section>
</section>
<section>
<section id="slide-org88a4369" data-background="img/al_scenarios.svg.p0.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org88a4369"></h2>
</section>
</section>
<section>
<section id="slide-org73234f4" data-background="img/al_scenarios.svg.p1.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org73234f4"></h2>
</section>
</section>
<section>
<section id="slide-orgc1b696c" data-background="img/al_scenarios.svg.p2.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="orgc1b696c"></h2>
</section>
</section>
<section>
<section id="slide-org6bec5a3" data-background="img/al_scenarios.svg.p3.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org6bec5a3"></h2>
</section>
</section>
<section>
<section id="slide-org38d9863" data-background="img/al_scenarios.svg.p4.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org38d9863"></h2>
</section>
</section>
<section>
<section id="slide-org1a3d323" data-background="img/al_scenarios.svg.p5.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org1a3d323"></h2>
</section>
</section>
<section>
<section id="slide-orge64496c" data-background="img/al_scenarios.svg.p6.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="orge64496c"></h2>
</section>
</section>
<section>
<section id="slide-org902ff6c" data-background="img/al_scenarios.svg.p7.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="org902ff6c"></h2>
</section>
</section>
<section>
<section id="slide-orgd49b3aa" data-background="img/al_scenarios.svg.p8.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="orgd49b3aa"></h2>
</section>
</section>
<section>
<section id="slide-orgb5b013b" data-background="img/al_scenarios.svg.p9.svg" data-background-size="50%" data-background-transition="slide">
<h2 id="orgb5b013b"></h2>
</section>
</section>
<section>
<section id="slide-orgc3b51dd">
<h2 id="orgc3b51dd">What is <b>Interesting</b>?</h2>
<ul class="fragment appear">
<li>uncertainty
<ul>
<li>least confident</li>
<li>margin</li>
<li>entropy</li>
</ul></li>
<li>query-by-committee</li>
<li>expected model change (decision theory)</li>
<li>expected error reduction</li>
<li>expected variance reduction</li>
<li>…</li>
</ul>
</section>
</section>
<section>
<section id="slide-orge7f6263" data-background="img/tcr.png" data-background-size="100%">
<h2 id="orge7f6263"></h2>
<aside class="notes">
<ul>
<li>active learning to make a robot explore and learn a room</li>
<li>learn properties of the world (movable and in what way)</li>
<li>expected information gain</li>
<li>robot created the data, not the human</li>
</ul>
</aside>
</section>
</section>
<section>
<section id="slide-orgb6b3bb2">
<h2 id="orgb6b3bb2">Gamification of Search</h2>
</section>
</section>
<section>
<section id="slide-org653900f" data-background="img/cars.png" data-background-size="100%">
<h2 id="org653900f"></h2>
</section>
</section>
<section>
<section id="slide-org54ffd4b" data-background="img/cars10.png" data-background-size="100%">
<h2 id="org54ffd4b"></h2>
</section>
</section>
<section>
<section id="slide-org7d51d42" data-background="img/cars11.png" data-background-size="100%">
<h2 id="org7d51d42"></h2>
</section>
</section>
<section>
<section id="slide-org1542f73" data-background="img/cars12.png" data-background-size="100%">
<h2 id="org1542f73"></h2>
</section>
</section>
<section>
<section id="slide-org61ed4be" data-background="img/cars13.png" data-background-size="100%">
<h2 id="org61ed4be"></h2>
</section>
</section>
<section>
<section id="slide-org82c6997" data-background="img/vegas.jpg" data-background-size="100%">
<h2 id="org82c6997"></h2>
</section>
<section id="slide-org13ee7ff">
<h3 id="org13ee7ff">Multi-Armed Bandits</h3>
<p>
Problem statement
</p>
<ol>
<li class="fragment roll-in">Find a multi-armed bandit</li>
<li class="fragment roll-in">Play arms using bandit theory</li>
<li class="fragment roll-in">Profit $$$</li>
</ol>
</section>
<section id="slide-org6cca3d8">
<h3 id="org6cca3d8">Problem statement</h3>
<ul>
<li>given a bandit with \(n\) arms</li>
<li>each arm \(i \in {1,\dots,n}\) returns reward</li>
</ul>
<p>
\[y \sim P(y; \theta_i)\]
</p>
<p class="fragment (roll-in)">
<b>Goal</b>: Find a policy that
\[\max \sum_{t=1}^T y_t\]
</p>
</section>
</section>
<section>
<section id="slide-orga1a0565">
<h2 id="orga1a0565">UCB</h2>
<p class="fragment roll-in">
past performance + exploration bonus
</p>
<aside class="notes">
<ul>
<li>Upper confident bound</li>
<li>Greedy</li>
<li>Not optimal, but bounded</li>
<li>"Optimism in the face of uncertainty"</li>
<li>Exploration vs Exploitation</li>
<li>Many variations of UCB</li>
</ul>
</aside>
</section>
<section id="slide-org6fa8660">
<h3 id="org6fa8660">UCB1</h3>
<p class="fragment roll-in">
Play each bandit once
</p>
<p class="fragment roll-in">
Then play bandit that \[\Large \arg\max_i \; \bar\mu_i + \sqrt{\frac{2\ln n}{n_i}}\]
</p>
<ul class="fragment roll-in">
<li>\(\bar\mu_i\): mean reward of bandit \(i\)</li>
<li>\(n\): total rounds played</li>
<li>\(n_i\): rounds bandit \(i\) was played</li>
</ul>
</section>
</section>
<section>
<section id="slide-org5ba9c3b">
<h2 id="org5ba9c3b">Demo</h2>
</section>
</section>
<section>
<section id="slide-orgeba45d7">
<h2 id="orgeba45d7">One Bandit per Feature</h2>
<ul>
<li class="fragment roll-in">brand bandit</li>
<li class="fragment roll-in">car body bandit</li>
<li class="fragment roll-in">segment bandit</li>
</ul>
<aside class="notes">
<ul>
<li>brand: Porsche, VW, …</li>
<li>car body: SUV, mini, copue, …</li>
<li>segment: sports car, economy car, mittelklasse, …</li>
</ul>
<p>
each bandit creates a <b>ranking</b> for the given feature
</p>
</aside>
</section>
</section>
<section>
<section id="slide-orgd44ae68">
<h2 id="orgd44ae68">Ranking with Elasticsearch</h2>
<aside class="notes">
<ul>
<li>made for creating rankings</li>
<li>output of bandits is input of elasticsearch query</li>
</ul>
</aside>
<p>
<img src="./img/es_ranking.png" alt="es_ranking.png" />
<img src="./img/es.png" alt="es.png" />
</p>
</section>
</section>
<section>
<section id="slide-org49125c4" data-background="img/bias.png" data-background-size="100%">
<h2 id="org49125c4">Popularity Bias</h2>
</section>
</section>
<section>
<section id="slide-org44e1300" data-background="img/segmentation.png" data-background-size="120%">
<h2 id="org44e1300"></h2>
<aside class="notes">
<ul>
<li>Sparse PCA to find set of sparse components that can optimally reconstruct the data</li>
<li>Then clustering</li>
</ul>
</aside>
</section>
</section>
<section>
<section id="slide-org757a326">
<h2 id="org757a326">Practical Remarks</h2>
<ul>
<li class="fragment roll-in">Python all the way down ;D</li>
<li class="fragment roll-in">sklearn</li>
<li class="fragment roll-in">Flask REST API</li>
<li class="fragment roll-in">Elasticsearch</li>
</ul>
</section>
</section>
<section>
<section id="slide-orgf4d82a3">
<h2 id="orgf4d82a3">Conclusion</h2>
<p class="fragment roll-in">
Active Learning or: How I Learned to Stop Worrying and Love Small Data
</p>
</section>
</section>
<section>
<section id="slide-orge3f2e0f">
<h2 id="orge3f2e0f">Related Topics</h2>
<ul>
<li>Sequential Decision Making</li>
<li>Global Optimizaiton</li>
<li>Experimental Design</li>
<li>(Bayesian) Reinforcement Learning</li>
<li>Optimal solution exists: planning in <b>belief space</b>, but is infeasible</li>
<li>Tuning hyperparams with <a href="https://github.com/zygmuntz/hyperband">Hyperband</a></li>
</ul>
</section>
</section>
<section>
<section id="slide-org139daf0">
<h2 id="org139daf0">Thanks!</h2>
<p>
<b>Questions?</b>
</p>
<p>
<i>Stefan Otte</i>
</p>
<p>
<a href="https://goo.gl/8JgirR">https://goo.gl/8JgirR</a>
</p>
<div class="figure">
<p><img src="./img/qr.png" alt="qr.png" />
</p>
</div>
</section>
</section>
<section>
<section id="slide-org6d93b75">
<h2 id="org6d93b75">References</h2>
<ul>
<li><a href="http://burrsettles.com/pub/settles.activelearning.pdf">Active Learning Literature Survey</a></li>
<li><a href="http://homes.dsi.unimi.it/~cesabian/Pubblicazioni/ml-02.pdf">Finite-time Analysis of the Multiarmed Bandit Problem - Auer et al</a></li>
<li><a href="https://ipvs.informatik.uni-stuttgart.de/mlr/marc/teaching/14-BanditsOptimizationActiveLearningBayesianRL.pdf">Bandits, Global Optimization, Active Learning, and Bayesian RL – understanding the common ground - Toussaint</a> <a href="https://www.youtube.com/watch?v=5rev-zVx1Ps">video</a></li>
</ul>
</section>
</section>
</div>
</div>
<script src="reveal.js/lib/js/head.min.js"></script>
<script src="reveal.js/js/reveal.js"></script>
<script>
// Full list of configuration options available here:
// https://github.com/hakimel/reveal.js#configuration
Reveal.initialize({
controls: true,
progress: true,
history: true,
center: true,
slideNumber: 'c',
rollingLinks: false,
keyboard: true,
overview: true,
theme: Reveal.getQueryHash().theme, // available themes are in /css/theme
transition: Reveal.getQueryHash().transition || 'slide', // default/cube/page/concave/zoom/linear/fade/none
transitionSpeed: 'default',
multiplex: {
secret: '', // null if client
id: '', // id, obtained from socket.io server
url: '' // Location of socket.io server
},
// Optional libraries used to extend on reveal.js
dependencies: [
{ src: 'reveal.js/lib/js/classList.js', condition: function() { return !document.body.classList; } },
{ src: 'reveal.js/plugin/markdown/marked.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: 'reveal.js/plugin/markdown/markdown.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: 'reveal.js/plugin/zoom-js/zoom.js', async: true, condition: function() { return !!document.body.classList; } },
{ src: 'reveal.js/plugin/notes/notes.js', async: true, condition: function() { return !!document.body.classList; } }]
});
</script>
</body>
</html>