-
Notifications
You must be signed in to change notification settings - Fork 8
/
mgc-simplex.html
580 lines (327 loc) · 13.1 KB
/
mgc-simplex.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
<!DOCTYPE html>
<html>
<head>
<title>MGC | SIMPLEX</title>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8"/>
<link rel="stylesheet" href="fonts/quadon/quadon.css">
<link rel="stylesheet" href="fonts/gentona/gentona.css">
<link rel="stylesheet" href="slides_style.css">
<script type="text/javascript" src="assets/plotly/plotly-latest.min.js"></script>
</head>
<body>
<textarea id="source">
class: left,
name:opening
## Multiscale Graph Correlation:
### A Knowledge Representation System for Discovering Latent Geometric Structure
PI: Joshua T. Vogelstein, Co-PIs: Carey E. Priebe, Cencheng Shen
<br><br><br>
<img src="images/funding/jhu_bme_blue.png" STYLE="HEIGHT:95px;"/>
<img src="images/funding/KNDI.png" STYLE="HEIGHT:95px;"/>
.foot[[[email protected]](mailto:jovo at jhu dot edu)
| <http://neurodata.io/talks/mgc-simplex.html>]
---
# Outline
- Theoretical Results
- Numerical Analysis
- Bleeding Edge
- Discussion
---
class: middle
# .center[Theoretical Results]
---
### Our Goal
Build a knowledge represetation system with the following properties:
1. can answer various kinds of queries
2. has a rigorous mathematical and statistical framework
3. enables subject matter experts to encode domain-specific information including context
--
Requires definitions for
1. query types to support, and
2. mathematical desiderata to satisfy.
---
### Queries of Interest
Given $N$ entities, each measured with up to $M$ modalities, <br>
as well as directly measuring $R$ relationships (eg, FB friends), <br>
discover and decipher relationships between subsets of:
1. entities,
2. modalities.
---
#### Mathematical and Statistical Desiderata
1. .r[Sufficiency]: Operate in any measurable metric spaces
2. .r[Uncertainty]: Represent and quantify uncertainty
3. .r[Complexity]: Can discover any complex relationship
4. .r[Consistency]: Sample estimator converges to truth
4. .r[Robustness]: Robust to model misspecifications
4. .r[Efficiency]: Computationally tractable/efficient
---
### Limitations of Existing Approaches
Every existing approach has severe limitations in $\geq$ 1 of these
1. .r[Sufficiency]: HHG, dCorr, HSIC
2. .r[Uncertainty]: Isomap, Diffusion Wavelets
3. .r[Complexity]: Kendall, Spearman
4. .r[Consistency]: Maximal Correlation, MIC
4. .r[Robustness]: Pearson
4. .r[Efficiency]: TDA
---
#### Our Key Principles
1. all knowledge from the data is encoded in sets of relationships
2. the optimal set size is a function of the (data,query) pair
---
#### Our Basic Strategy
Given $N$ entities, each measured with up to $M$ modalities, <br>
as well as directly measuring $R$ relationships, <br>
build the following data structures:
1. For each modality
1. choose a metric
1. build the k-nearest neighbor graph, for all k=1,...,$N$
2. build a (centered) pairwise distance matrix
1. For each relationship
1. build the k-nearest neighbor graph, for all k=1,...,$N$
2. convert the relationship into a centered distance matrix
3. Estimate the jointly optimal set of neighborhood sizes simultaneously across all modalities
3. For each modality, sparsify/truncate/censor its pairwise distance matrix to only include the those neighbors
4. Compute the Hadamard tensor product
---
### Multiscale Graph Correlation
<img src="images/linear.png" alt="Drawing" style="position:absolute; TOP: 150px; LEFT: 100px; height: 180px;"/>
--
<img src="images/Fig1B.png" alt="Drawing" style="position:absolute; TOP: 150px; LEFT: 300px; height: 180px;"/>
<!-- --
<img src="images/Fig_centered1Panel2.png" alt="Drawing" style="position:absolute; TOP: 150px; LEFT: 500px; height: 150px;"/> -->
--
<div style="position:absolute; TOP: 150px; LEFT: 560px; height: 200px;">
dcorr(X,Y)=0.15, p-val < 0.001 <br> MGC(X,Y)=0.15, p-val < 0.001
</div>
--
<img src="images/spiral.png" alt="Drawing" style="position:absolute; TOP: 400px; LEFT: 100px; height: 180px;"/>
--
<img src="images/Fig8B.png" alt="Drawing" style="position:absolute; TOP: 400px; LEFT: 300px; height: 180px;"/>
<!-- --
<img src="images/Fig_centered8Panel2.png" alt="Drawing" style="position:absolute; TOP: 400px; LEFT: 500px; height: 150px;"/> -->
--
<div style="position:absolute; TOP: 400px; LEFT: 560px; height: 200px;">
dcorr(X,Y)=0.01, p-val 0.3 <br> MGC(X,Y)= .r[0.13], p-val < .r[0.001]
</div>
---
### Theoretical Background
.r[Non-Parametric Dependence] is the primary quantity of interest:
- bounds predictive accuracy
- bounds causal inference
- used for feature selection / screening
- mutual information is a special case
---
### Theoretical Results
- (nearly) all hypotheses can be stated in terms of dependence
- MGC .r[connects] previously disparate data science approaches
- kernel learning
- distance-based
- nearest-neighbors
- multiscale analysis
---
### Theoretical Limitations
MGC theoretically satisfies all of the desiderata, with the following caveats, MGC requires
1. .r[Sufficiency]: metric must be of "strong negative type"
2. .r[Uncertainty]: samples are "exchangable" with finite variance
3. .r[Complexity]: $R=2$ or $M=2$
4. .r[Consistency]: $\mathcal{O}(1/n)$
4. .r[Robustness]: non-parametric
4. .r[Efficiency]: $\mathcal{O}(n^2 \log n)$
---
class: middle
# .center[Numerical Analysis]
---
class: top, left
### Setup
- sample $(x_i, y_i) \sim F$ iid for $i=1,\ldots, n$ for 20 different functions
- in each case, there is *some* true but unknown relationship
- *power* is the probability detecting a relationship
<!-- - $\beta_n(t)$ := power of test statistic $t$ given $n$ samples -->
- $N_\beta$ is the # of samples required to achieve power $\beta$
---
class: top, left
### 20 Different Functions (1D version)
<img src="images/FigSimVisual.png" alt="local" style="width: 100%;"/>
---
class: top, left
### MGC Outperforms Benchmarks
<img src="images/FigHDPower.png" alt="local" style="width: 100%;"/>
<!-- -- -->
<!-- <img src="images/FigHDPowerSummarySize.png" alt="local" style="position:absolute; LEFT: 460px; height: 280px;"/> -->
<!-- .center[
<img src="https://github.com/neurodata/MGC/raw/master/Figures/FigHDPowerMGCM.png" alt="Drawing" style="width: 900px;"/>
] -->
---
### MGC Outperforms Benchmarks
Relative sample size to achieve 85% power
<img src="images/MGC_power_table.png" alt="local" style="width: 100%;"/>
Other approaches with the sample theoretical properties and limitations (HHG, dCorr, HSIC)
tend to require **double** or **triple** the samples to achieve the same power
---
### MGC Biomedical Applications
- detecting dependence (human brain imaging)
- graph topology vs node attributes (worm network)
- signal subgraph detection (human brain network and personality)
- effect size estimation ("batches" of imaging)
- feature selection/screening (cancer genetics)
- machine learning / classification (proteomics)
- DARPA's SIGMA data
---
class: middle
# .center[Bleeding Edge]
---
### DARPA's Sigma Data
- Details of the analysis will be provided after lunch
- Existing MGC theory and methods were inadequate to perform the tasks
- Computationally, MGC required $\mathcal{O}(r \times n^2 \log n)$ run time,
where $r$ is the number of permutations required in the permutaiton test
- SIGMA data had $n=25,000$, which was too large for existing MGC to handle
---
### Accelerated MGC
- Goal: accelerate MGC without weakening theory
- Reduce to linear time without permutation: $\mathcal{O}(r \times n^2 \log n) \rightarrow \mathcal{O}(n \log n)$
- Algorithm:
- For $i = 1, \ldots, \sqrt{N}$
- randomly subsample $\sqrt{n}$ points
- apply MGC to obtain $t_i$
- compute $\langle t_i \rangle$
- Variance of sub-sample provides confidence intervals
- Reduces computational time in practice from hours to minutes
- But minor loss of power
---
### Extended Results
1. Proved equivalence of existing kernel and distance approaches
2. Proved random forest (RF) induces a universal kernel
3. Demonstrated empirically that RF yields improved results
---
### Next Steps
By coupling RF with MGC, we strengthen the theoretical results
1. Sufficiency: metric must be of "strong negative type"
- .r[RF can learn metric]
2. Uncertainty: samples are "exchangable" with finite variance
- .r[just requires exchangability]
3. Complexity: $R=2$ or $M=2$
- .r[RF can handle $R>2$]
4. Consistency: $\mathcal{O}(1/n)$
- .r[$\mathcal{O}(1/n^2)$?]
4. Robustness: non-parametric
- .r[yes]
4. Efficiency: $\mathcal{O}(n^2 \log n)$
- .r[$\mathcal{O}(n \log n)$]
And RF adds additional queries, including entropy, conditional entropy, mutual information, classification, and regression
---
class: middle
# .center[Discussion]
---
### References
- MGC foundational theory [[1]](https://arxiv.org/abs/1710.09768)
- MGC biomedical applications [[2]](https://arxiv.org/abs/1609.05148)
- MGC for independence between graph topology & attributes [[3]](https://arxiv.org/abs/1703.10136)
- MGC for signal subgraph detection [[4]](https://arxiv.org/abs/1801.07683)
- MGC for clustering (ish) [[5]](https://arxiv.org/abs/1710.09859)
- Equivalence of distance and kernel learning [[6]](https://arxiv.org/abs/1806.05514)
- R package: [CRAN link](https://cran.r-project.org/web/packages/mgc/index.html); ~80 monthly downloads
---
### Summary
- MGC is a KRS that can discover latent geometric structure
- We have used it to answer various query types
- Has a rigorous mathematical and statistical foundation
- SME can encode domain specific knowledge
- Coupling to RF generalizes, improves theoretical and empirical performances
- Specifically useful for DARPA SIGMA (as we will show later)
---
class: top, left
### Acknowledgements
<div class="container">
<img src="faces/cep.png"/>
<div class="centered">Carey Priebe</div>
</div>
<div class="container">
<img src="faces/cshen.jpg"/>
<div class="centered">Cencheng Shen</div>
</div>
<div class="container">
<img src="faces/shangsi.jpg"/>
<div class="centered">Shangsi Wang</div>
</div>
<div class="container">
<img src="faces/youjin-lee.jpg"/>
<div class="centered">Youjin Lee</div>
</div>
<div class="container">
<img src="faces/ebridge.jpg"/>
<div class="centered">Eric Bridgeford</div>
</div>
<span style="font-size:200%; color:red;">♥, 🦁, 👪, 🌎, 🌌</span>
<img src="images/funding/nsf_fpo.png" STYLE="HEIGHT:95px;"/>
<img src="images/funding/nih_fpo.png" STYLE="HEIGHT:95px;"/>
<img src="images/funding/darpa_fpo.png" STYLE=" HEIGHT:95px;"/>
<img src="images/funding/iarpa_fpo.jpg" STYLE="HEIGHT:95px;"/>
<img src="images/funding/KAVLI.jpg" STYLE="HEIGHT:95px;"/>
<img src="images/funding/schmidt.jpg" STYLE="HEIGHT:95px;"/>
---
### Questions?
<img src="images/liono.JPG" STYLE="position:absolute; TOP:110px; HEIGHT:500px;"/>
---
### DARPA SIGMA Data
Some radioactive substance is moving around, and we have
- 3 days of data
- 25 sensors
- 15 features per sensor
- each feature sampled at different frequency
- some features on some days include $\approx$ 25,000 samples
- 2 features of interest of particular: gamma and neutron
--
### Queries
1. Do any sensors have information about radiation? If so, which?
2. Do any features have information about radiation? If so, which?
3. Are any features redundant?
---
### Data Processing Pipeline
1. Open hdf5 files
2. Convert to 2D arrays
3. Run MGC on each
4. Visualize results
---
### What does raw data look like?
<img src="images/timevsAltitude.jpg" alt="local" style="width: 100%;"/>
---
### Do any sensors/features have information about gamma?
<img src="images/featurevsgamma.jpg" alt="local" style="width: 100%;"/>
---
### Do any sensors/features have information about neutron?
<img src="images/featurevsneutron.jpg" alt="local" style="width: 100%;"/>
<!-- background-image: url("images/OCPaper.png")
background-size: contain;
-->
<!-- <img src="images/OCPaper.png" style="width: 100%;"/> -->
---
### Are any features correlated with one another?
<img src="images/pvalAverage.jpg" alt="local" style="width: 100%;"/>
---
### SIGMA Conclusions
1. Neutron measurements seem to be redundant
2. Several of the features are uninformative with regard to radiation (eg, location, voltage, etc.)
3. All features that are informative about Gamma are also informative about one another
</textarea>
<!-- <script src="https://gnab.github.io/remark/downloads/remark-latest.min.js"></script> -->
<script src="remark-latest.min.js"></script>
<script src="assets/KaTeX/0.5.1/katex.min.js"></script>
<script src="assets/KaTeX/0.5.1/auto-render.min.js"></script>
<link rel="stylesheet" href="assets/KaTeX/0.5.1/katex.min.css">
<script type="text/javascript">
var options = {};
var renderMath = function() {
renderMathInElement(document.body);
// or if you want to use $...$ for math,
renderMathInElement(document.body, {delimiters: [ // mind the order of delimiters(!?)
{left: "$$", right: "$$", display: true},
{left: "$", right: "$", display: false},
{left: "\\[", right: "\\]", display: true},
{left: "\\(", right: "\\)", display: false},
]});
}
var slideshow = remark.create(options, renderMath);
</script>
</body>
</html>