/
2017_Seminar.html
469 lines (412 loc) · 17.6 KB
/
2017_Seminar.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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<title>Final Thesis Seminar (2017)</title>
<meta name="description" content="Final Thesis Seminar (2017)">
<meta name="author" content="Philippe Desjardins-Proulx">
<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">
<link rel="stylesheet" href="css/reveal.min.css">
<link rel="stylesheet" href="css/theme/polaris.css" id="theme">
<link rel="stylesheet" href="lib/css/zenburn.css">
<script>
if (window.location.search.match(/print-pdf/gi)) {
var link = document.createElement('link');
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = 'css/print/pdf.css';
document.getElementsByTagName('head')[0].appendChild(link);
}
</script>
<!--[if lt IE 9]>
<script src="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>Automatic Revision of Ecological Theories</h1>
<p>by <a href="http://phdp.github.io/">Philippe Desjardins-Proulx</a></small></p>
</section>
<section>
<section>
<h1>Motivation</h1>
</section>
<section>
<img src='img/amnat12.png' alt='amnat12'/>
</section>
<section>
<img src='img/eco-evo-paper.png' alt='ecoevo'/>
</section>
<section>
<img src='img/complex-1.svg' width='70%' alt='complex fig 1'/>
</section>
<section>
<img src='img/complex-0.svg' width='100%' alt='complex fig 0'/>
</section>
<section>
<img src='img/wagner.png' alt='wagner'/>
</section>
<section>
<img src='img/hahn08.png' alt='hahn08'/>
</section>
<section>
<h2>[Supervised] learning in a nutshell</h2>
<img src='img/ml-900.png' alt='machine learning'/>
</section>
<section>
<h2>Random forests</h2>
<img src='img/random_forests.png' alt='machine learning'/>
</section>
<section>
<img src='img/biorxiv-scientifictheories.png' alt='biorxiv'/>
</section>
<section>
<h2>The thesis' thesis</h2>
<p>Biodiversity is shaped by too many factors (speciation,
stochastic drift, selection from abiotic and biotic factors) to be
understood with a single unifying theory.</p>
<br/>
<p>Thus, we must combine traditional theories with techniques
capable of learning from data, but we must be able to keep the
structure and clarity of these traditional theories.</p>
</section>
<section>
<img src='img/reasoning_systems.png' alt='reasoning systems'/>
</section>
<section>
<h2>Lotka-Volterra framework</h2>
\[\frac{dx}{dt} = \alpha x - \beta xy \mbox{ and } \frac{dy}{dt} = \delta xy - \gamma y.\]
</section>
<section>
<h2>Probabilistic niche model</h2>
\[ PPreyOn(x, y) = \alpha \times \exp\left[-\left(\frac{N(y) - D(x)}{R(x)/2}\right)^2\right].\]
</section>
<section>
<h2>Exponential growth</h2>
<p>From</p>
\[N(x, t + 1) = G(x) \times N(x, t).\]
<br/>
<p>To</p>
\[SmallN(x, t) \mbox{ and } Resources(x, t) \Rightarrow N(x, t + 1) = G(x) \times N(x, t).\]
</section>
<section>
<h2>Theory revision wishlist</h2>
<p>Handling deterministic and probabilistic theories.</p>
<p>Fast algorithms for theory revision.</p>
<p>Handling arbitrarily large knowledge bases.</p>
</section>
</section>
<section>
<section>
<h1>Ecological Interactions</h1>
</section>
<section>
<h2>Why interactions?</h2>
<p>First reason: a key reason why selection is difficult to handle
with traditional theories is because of the complexity of species
interactions.</p>
<br/>
<p>Second reason: my lab is studying interactions. I had access to
both good data-sets and people who understood them well.</p>
</section>
<section>
<img src='img/ecological-netflix-paper.png' alt='paper'/>
</section>
<section>
<h2>Data</h2>
<ul width='100%'>
<li>Assembled by Isabelle Daigle, mostly from Digel et al. (2014).</li>
<li>48 soil food webs.</li>
<li>881 species.</li>
<li>34 193 unique interactions.</li>
<li>215 418 absence of interactions.</li>
<li>3 real valued-traits: \(body mass\), \(Ph0\), \(Ph1\).</li>
<li>25 binary traits (e.g. AboveGround, Detritivore, Bacteria, Jumps, LongLegs, UsePoison).</li>
</ul>
</section>
<section>
<h2>K nearest neighbour (KNN) with K = 3</h2>
<img src='img/pearson.svg' width='100%' alt='pearson'/>
</section>
<section>
\[tanimoto(\mathbf{x}, \mathbf{y}) = \frac{\left\vert\mathbf{x} \cap \mathbf{y}\right\vert}{\left\vert\mathbf{x} \cup \mathbf{y}\right\vert}.\]
<br/><br/>
\begin{align}
tanimoto(\mathbf{x}, \mathbf{y}) &= \frac{\left\vert\{A, B, C\} \cap \{A, C, D, E, F\}\right\vert}{\left\vert\{A, B, C\} \cup \{A, C, D, E, F\}\right\vert}\\
&= \frac{\left\vert\{A, C\}\right\vert}{\left\vert\{A, B, C, D, E, F\}\right\vert} = \frac{2}{6}.
\end{align}
</section>
<section>
\[TSS = \frac{(tp \times tn) - (fp \times fn)}{(tp + fn)(fp + tn)}.\]
<br/><br/>
<p><i>tp, tn, fp, fn</i> are the true positives, true negatives, false positives,
and false negaties, respectively.</p>
<p>The <i>TSS</i> ranges from -1 to 1.</p>
</section>
<section>
<p>Supervised learning Random forests predict correctly 99.55% of
the non-interactions and 96.81% of the interactions, for a
<i>TSS</i> of 0.96.</p>
<br/>
<p>Removing the binary traits has little effect on the model. With
only \(body mass\), \(Ph0\), \(Ph1\), the <i>TSS</i> of the random
forest is 0.94.</p>
</section>
</section>
<section>
<section>
<h1>KNN with Threshold</h1>
</section>
<section>
<img src='img/knn-paper.png' alt='knn'/>
</section>
<section>
<h2>I Bartomeus' pollinator-plant data-set</h2>
<ul width='100%'>
<li>Plant-pollinator interactions were collected across 16 sites in
the southwest of the iberian peninsula (Huelva and Seville).</li>
<li>65 pollinators.</li>
<li>277 plants.</li>
<li>739 pollinator-plant interactions.</li>
<li>We randomly pick 739 pollinator-plant pairs that are not
interacting to build a data-set of 1478 entries, half of
interactions, half of non-interactions.</li>
</ul>
</section>
<section>
<h2>Accuracy</h2>
<img src='img/ht_acc.svg' width='80%' alt='acc'/>
</section>
<section>
<h2>TSS</h2>
<img src='img/ht_tss.png' width='80%' alt='acc'/>
</section>
</section>
<section>
<section>
<h1>Markov Logic</h1>
</section>
<section>
\[\mbox{Forall } x, y, z: Friends(x, y) \mbox{ and } Friends(y, z) \Rightarrow Friends(x, z), 0.7.\]
\[\mbox{Forall } x: Smoking(x) \Rightarrow Cancer(x), 1.5.\]
\[\mbox{Forall } x, y: Friends(x, y) \mbox{ and } Smoking(x) \Rightarrow Smoking(y), 1.1.\]
<br/><br/>
<p>Can answer queries of the form:</p>
<br/>
\[P(Cancer(Bob) | Smoking(Anna) \mbox{ and } Friends(Anna, Bob)).\]
\[P(Friends(x, y) \Leftrightarrow Friends(y, x) | ...).\]
</section>
<section>
<h2>Salix data-set</h2>
<img src='img/salix.png' alt='salix'/>
\[Parasites \rightarrow Galler \rightarrow Salix\]
</section>
<section>
<h2>Learning weights (but not really rules)</h2>
\[HighNumSalix(x), 6.44.\]
\[\neg ModerateCooccurrence(x,y), 5.60.\]
\[\neg PreyOnAt(x,y,a3), 4.38.\]
\[\neg HighCooccurrence(x,y), 4.33.\]
\[\neg PresenceAt(x,y), 3.08.\]
\[\neg PreyOn(x,y), 1.68.\]
\[\neg IsSalix(x), 1.38.\]
\[\neg IsGaller(x), 0.59.\]
\[\neg CloselyRelated(x,y), 0.59.\]
\[\neg HigherPhyloValue(x,y), 0.57.\]
</section>
<section>
\[ HighNumGaller(x), 0.47\]
\[ HighNumSpecies(x), 0.24.\]
\[ HighFoodWebLinkDensity(x), 0.19.\]
\[ \neg IsParasitoid(x), 0.18.\]
\[ \neg HighFoodWebConnectance(x), 0.14.\]
\[ \neg HighTotalPrecipiration(x), 0.01.\]
\[ \neg HighMeanTemperature(x), 0.01.\]
\[ \neg PreyOn(x,x), 0.01.\]
\[ HighNumParasites(x), 0.00.\]
</section>
<section>
\[ PreyOnAt(x,y,z) \Rightarrow PreyOn(x,y), 5.20.\]
\[ HighCooccurrence(x,y) \Rightarrow PreyOn(x,y), 4.22.\]
\[ IsGaller(x) \mbox{ and } PreyOn(x,y) \Rightarrow IsSalix(y), 4.15.\]
\[ IsParasitoid(y) \mbox{ and } PreyOn(y,x) \Rightarrow IsGaller(x), 3.49.\]
\[ PreyOn(x,y) \Rightarrow HighCooccurrence(x,y), 1.57.\]
\[ PreyOn(x,y) \Rightarrow PreyOn(x,x), 1.52.\]
\[ HigherPhyloValue(x,y) \Rightarrow PreyOn(x,y), 1.13.\]
\[ HigherPhyloValue(x,y) \Rightarrow HighCooccurrence(x,y), -0.59.\]
\[ PreyOn(x,x) \Rightarrow PreyOn(x,y), 0.45.\]
\[ \neg IsSalix(x) \mbox{ or } \neg PreyOn(x,y), 0.02.\]
\[ CloselyRelated(x,y) \mbox{ and } PreyOn(x,z) \Rightarrow PreyOn(y,z), 0.00.\]
</section>
</section>
<section>
<section>
<h1>Fuzzy Logic</h1>
</section>
<section>
<h2>I Bartomeus' pollinator-plant data-set (part II)</h2>
<p>Seven real-valued traits:</p>
<ul>
<li>Plant's nectar tube dimension.</li>
<li>Plant's nectar tube depth.</li>
<li>Plant's flower width.</li>
<li>Pollinator's body size.</li>
<li>Pollinator's tongue length.</li>
<li>Pollinator generalism.</li>
<li>\(K\) nearest neighbour ratio.</li>
</ul>
</section>
<section>
<h2>Structure</h2>
\[\mathbf{If} \mbox{ } antecedants \mbox{ } \mathbf{then} \mbox { } consequents.\]
\[\mathbf{If} \mbox{ } antecedants \mbox{ } \mathbf{then} \mbox { } consequents.\]
\[\mathbf{If} \mbox{ } antecedants \mbox{ } \mathbf{then} \mbox { } consequents.\]
\[...\]
<br/>
<br/>
\[\mathbf{If} \mbox{ } A_0 \mbox{ and } A_1 \mbox{ and } A_2 \mbox{ and } ... \mbox{ } \mathbf{then} \mbox { } C_0.\]
</section>
<section>
<h2>Initial model</h2>
<ul style='width: 100%'>
<li style='color:#246325'><b>If</b> KNN is high <b>then</b> interaction.</li>
<li style='color:#632427'><b>If</b> KNN is low <b>then</b> non-interaction.</li>
</ul>
</section>
<section>
<h2>Current approach for learning (Python script)</h2>
<h3>Evolution</h3>
<ol start='0'>
<li>Start with the initial knowledge base with the two KNN rules.</li>
<li>Randomly generate a rule, add it to the knowledge base.</li>
<li>If the new knowledge base beats the old by 0.004 TSS, keep the rule.</li>
<li>Repeat \(\phi\) times</li>
</ol>
<br/><br/>
<h3>Pseudo-ensemble approach</h3>
<ol start='0'>
<li>Make \(n\) evolutions.</li>
<li>Count the frequency of the rules in the evolutions</li>
<li>Try to add the rule one by one from most frequent to least</li>
<li>Keep if TSS improves by at least 0.004.</li>
</ol>
</section>
<section>
<ul style='width: 100%; font-size: 80%'>
<li style='color:#246325'><b>If</b> Plant nectar tube depth is average <b>then</b> interaction.</li>
<li style='color:#246325'><b>If</b> Pollinator body size is small <b>then</b> interaction.</li>
<li style='color:#246325'><b>If</b> KNN is high <b>then</b> interaction.</li>
<li style='color:#246325'><b>If</b> Pollinator is specialist <b>and</b> Plant nectar tube dim is small <b>then</b> interaction.</li>
<li style='color:#246325'><b>If</b> Plant flower width is average <b>then</b> interaction.</li>
<li style='color:#246325'><b>If</b> Plant flower width is small <b>and</b> Pollinator tongue length is small <b>then</b> interaction.</li>
<li style='color:#632427'><b>If</b> KNN is low <b>then</b> non-interaction.</li>
</ul>
</section>
<section>
<h2>Results on pollinator-plant data-set</h2>
<table style="width:100%; text-align:center;">
<tr>
<th>Model</th>
<th>TSS (10% testing data)</th>
</tr>
<tr>
<td>Support Vector Machine</td>
<td>0.564</td>
</tr>
<tr>
<td>Fuzzy (initial model with two KNN rules)</td>
<td>0.689</td>
</tr>
<tr>
<td>Decision tree</td>
<td>0.723</td>
</tr>
<tr>
<td>Random forests</td>
<td>0.795</td>
</tr>
<tr>
<td>Fuzzy (after theory revision)</td>
<td>0.841</td>
</tr>
</table>
</section>
</section>
<section>
<section>
<h1>Conclusions</h1>
</section>
<section>
<ul width='100%'>
<li>KNN is an effective, widely applicable, and simple approach.</li>
<li>KNN with threshold works well with TSS scores.</li>
<li>Supervised learning algorithms (esp. random forests) can predict interactions with few real-valued traits.</li>
<li>Fuzzy knowledge bases are both clear and accurate.</li>
<li>Even a silly revision algorithm yields very good results for fuzzy logic.</li>
<li>Knowledge representation in ecology is an interesting topic in itself.</li>
<li>Markov logic has issues (computational, binary) but modern variants could be promising.</li>
</ul>
</section>
</section>
<section>
<section>
<h1>What's Next</h1>
</section>
<section>
<h2>Salix part II?</h2>
<img src='./img/psl_paper.png' alt='psl paper'/>
</section>
<section>
<h2>...also fuzzy logic meets salix</h2>
</section>
<section>
<h2>Genetic (or evolutionary) algorithms</h2>
<img src='./img/generic-algorithms.png' alt='evolutionary algorithms'/>
</section>
</section>
<section>
<section>
<h1>Questions?</h1>
</section>
</section>
</div>
</div>
<script src="lib/js/head.min.js"></script>
<script src="js/reveal.min.js"></script>
<script>
// Full list of configuration options available here:
// https://github.com/hakimel/reveal.js#configuration
Reveal.initialize({
width: "70%",
height: "100%",
margin: 0,
minScale: 1,
maxScale: 1,
controls: true,
progress: true,
history: true,
center: true,
theme: Reveal.getQueryHash().theme, // available themes are in /css/theme
transition: Reveal.getQueryHash().transition || 'linear', // default/cube/page/concave/zoom/linear/fade/none
// Parallax scrolling
// parallaxBackgroundImage: 'https://s3.amazonaws.com/hakim-static/reveal-js/reveal-parallax-1.jpg',
// parallaxBackgroundSize: '2100px 900px',
// Optional libraries used to extend on reveal.js
dependencies: [
{ src: 'lib/js/classList.js', condition: function() { return !document.body.classList; } },
{ src: 'plugin/markdown/marked.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: 'plugin/markdown/markdown.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: 'plugin/highlight/highlight.js', async: true, callback: function() { hljs.initHighlightingOnLoad(); } },
{ src: 'plugin/zoom-js/zoom.js', async: true, condition: function() { return !!document.body.classList; } },
{ src: 'plugin/notes/notes.js', async: true, condition: function() { return !!document.body.classList; } }
]
});
</script>
<script type="text/javascript" src="../../js/bower/MathJax/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
</body>
</html>