-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathL2F_6mo.html
402 lines (243 loc) · 10 KB
/
L2F_6mo.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
<!DOCTYPE html>
<html>
<head>
<title>L2F | L2M</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
## Lifelong Learning Forests
PI: Joshua T. Vogelstein, Co-PIs: Carey E. Priebe, Randal Burns
<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/L2F_6mo.html>]
---
# Outline
- Background
- Methods Update
- Applications Update
- Discussion
---
# Background
---
### Decision Forests Empirically Rock!
- Caruana et al. 2006 (ICML): "With excellent performance on all eight metrics, calibrated .r[boosted trees] were the best learning algorithm overall. .r[Random forests] are close second."
- Caruana et al. 2008 (ICML): "the method that performs consistently well across all dimensions is .r[random forests]."
- Delgado et al. 2014 (JMLR): "The classifiers most likely to be the bests are the .r[random forest]."
- Chen et al. 2016 (KDD): "Among the 29 challenge winning solutions published at Kaggle’s blog during 2015, 17 solutions used .r[XGBoost]. Among these solutions, eight solely used .r[XGBoost] to train the model"
---
### Decision Forests Theoretically Rock!
2. **Consistency** under certain assumptions
3. **Efficiency** error scales with order 1/n
1. **Uncertainty** provide consistent estimate of posteriors
4. **Flexibility** can learn simple or complex functions
4. **Stability** are robust
5. **Scalability** linear in n, p, and 1/T
6. **Explainability** feature importance available
7. **Automaticity** very few parameters to tweak
---
### LumberJack (LJ)
<img src="images/rerf_perf.png" style="width: 800px;"/>
- previously known as [Randomer Forest](https://arxiv.org/abs/1506.03410)
- replace feature selection with very sparse random projection
- significant empirical improvement over RF/XGBboost
- maintains all 8 theoretical properties above
- but are not lifelong learners
---
### Lifelong Learning Forests (L2F)
Extend LumberJack to maintain all the above properties,
and continue improving empirically even after things change (with or without labels).
#### Route
1. Prove Universal Consistency of LJ (with empirical support)
2. Prove Universal Kernel induced by LJ (with empirical support)
3. Prove LJ Kernel is Robust to noise dimensions (and therefore, changes in them, with empirical support)
4. Prove LJ Kernel is updated when signal changes too (rest of Phase I)
---
### Quick Primer on Decision Forests
- given n samples from data in p dimensions
- for each tree
- subsample m samples
- for each leaf node
- randomly sample/generate p features
- choose the feature/split the best "fits" the data
- step when further splitting does not help
---
### Result 1: *Universal* Consistency of LJ
<img src="images/error_rate_consistency.png" style="width: 800px;"/>
Conjecture 1: Lumberjack is at least as consistent as RF.
--
"Thm" 1: LJ is *more* consistent than RF.
---
### Result 2: Induced *Universal* Kernel
- Universal kernels are foundational in machine learning (e.g., Boltzman Machines, Deep Boltzman Machines)
- Existing kernels are pre-defined / inflexible
--
- LJ induced kernel: the proximity between a pair of points is the fraction of trees for which that pair lies in the same terminal node.
- Conjecture 2: the LJ induced kernel is universal.
--
"Thm" 2: The kernel induced by RF/LJ can trivially be modified to be universal.
---
#### Learned Kernel > "Pre-Specified" Kernels
<img src="images/rf_kernel.png" style="width: 700px;"/>
---
#### Result 3: LJ Kernel is Robust to High-Dimensional Noise
---
### Hardy Weinberg Curve
- $t_i \sim Beta(\alpha,\beta$)
- $x_i = [(t_i)^2, (1- t_i) t_i, (1-t_i)^2]$
- 3D embedding of CMDS recovers latent manifold
<img src="images/EUC_DIS_CMDS_d=0.png" style="height: 400px;"/>
---
### Hardy Weinberg Curve with Noise
- $x_i \leftarrow [x_i,$ 17 dimensions of noise]
- 3D embedding of CMDS recovers all .r[noise]
<img src="images/EUC_DIS_CMDS_d=17.png" style="height: 400px;"/>
---
### Hardy Weinberg Curve with Noise
- $x_i \leftarrow [x_i,$ 17 dimensions of noise]
- 3D embedding of RF learned kernel recovers all ".r[manifold]"
<img src="images/projected-normalnoise-hw-data-500trees-2000pts-17noisedims-depth4.jpeg" style="height: 400px;"/>
---
### Noise Dimensions Theorem
"Thm" 3: Nearest neighbors are robust to added noise dimensions
---
### Summary of Methods Update
- L2M with theoretical guarantees requires:
- Universal Consistency
- A Kernel that learns
- Learned Kernel must be universal (or nearly)
- We showed
- Lumberjack has all 3 of those
- Empirical ramifications even for static learning
---
# Applications
---
### \#1: Characterizing Psychopathy with Changing Sensors
- 10 years of scanning psychopaths (multi-modality)
- Frequent scanner updates meticulously documented (changing sensors)
- Task: Estimate recidivism using the first n samples, then improve accuracy including m additional unlabeled samples
- Status:
- 4,000+ individuals so far
- All participants have detailed clinical assessments (i.e., mental health, IQ, neuropsych)
- All participants complete structural (T1/T2-FLAIR), Diffusion, Resting state and fMRI
- Detailed outcomes data being collected to examine risk for crime (property, nonviolent, violent, sexually violent).
- Data organized and ready to be processed
- Data wrangling code is ready to run
---
<img src="images/MRN_bus.png" style="position:absolute; left:-2px; width:920px;"/>
---
### \#2: Characterizing Personality in Non-Stationary Life Span Data
- Lifespan data: 9-90
- Certain phenotypics are fixed (e.g., ethnicity, gender, IQ)
- Multimodal brain measurements are dynamic with age
- Task: Predict phenotypic properties using pediatric data, then improve accuracy using geriatric data
- Status:
- 180 individuals scanned
- 171 processed
- basic demographics available for each
- age, sex, handedness, race, etc.
- predictions are ready to be made
---
### Next Steps
- Complete wrangling of all image and phenotypic data
- Train LJ on "early" data
- Continue learning using additional data
- Benchmark performance vs. using full data
---
<div class="background">
<img src="images/lion_l2m.JPG" STYLE="position:absolute; TOP:0px; LEFT: 0px; HEIGHT:680px; WIDTH=600px;">
<div class="bottom-center"> <h1>#3: Lion</p></h1> </div>
</div>
---
### References
- Lumberjack [[1]](https://arxiv.org/abs/1506.03410)
- The Exact Equivalence of Distance and Kernel Methods for Hypothesis Testing [[2]](https://arxiv.org/abs/1806.05514)
- Forest Packing: Fast, Parallel Decision Forests [[3]](https://arxiv.org/abs/1806.07300)
- Kernel k-Groups via Hartigan's Method [[4]](https://arxiv.org/abs/1710.09859)
- From Distance Correlation to Multiscale Graph Correlation [[5]](https://arxiv.org/abs/1710.09768) (in press at JASA)
- Lumberjack R package: [CRAN](https://cran.r-project.org/web/packages/rerf/index.html)
---
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/randal.jpg"/>
<div class="centered">Randal Burns</div>
</div>
<div class="container">
<img src="faces/cshen.jpg"/>
<div class="centered">Cencheng Shen</div>
</div>
<div class="container">
<img src="faces/minh.jpg"/>
<div class="centered">Minh Tang</div>
</div>
<div class="container">
<img src="faces/percy.jpg"/>
<div class="centered">Percy Li</div>
</div>
<div class="container">
<img src="faces/tyler.jpg"/>
<div class="centered">Tyler Tomita</div>
</div>
<div class="container">
<img src="faces/james.jpg"/>
<div class="centered">James Browne</div>
</div>
<div class="container">
<img src="faces/falk_ben.jpg"/>
<div class="centered">Ben Falk</div>
</div>
<div class="container">
<img src="faces/jesse.jpg"/>
<div class="centered">Jesse Patsolic</div>
</div>
<div class="container">
<img src="faces/loftus.jpg"/>
<div class="centered">Alex Loftus</div>
</div>
<div class="container">
<img src="faces/veronika.jpg"/>
<div class="centered">Veronika Strnadova</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?
</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>