-
Notifications
You must be signed in to change notification settings - Fork 8
/
DARPA_L2M_PI_2022.html
594 lines (423 loc) · 16.5 KB
/
DARPA_L2M_PI_2022.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
<!DOCTYPE html>
<html>
<head>
<title>Learning</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_v2.css"> -->
<link rel="stylesheet" href="slides_style_i.css">
<script type="text/javascript" src="assets/plotly/plotly-latest.min.js"></script>
</head>
<body>
<textarea id="source">
# Lifelong Learning:<br>What is it anyway?
<br>
PI: Joshua T. Vogelstein, [JHU](https://www.jhu.edu/) <br>
Co-PI: Vova Braverman, [JHU](https://www.jhu.edu/) <br>
Jayanta Dey, Ali Geisa, Will LeVine, Hayden Helm, Ronak Mehta, Haoran Li, Aditya Krishnan, Rahul Ramesh, Pratik Chaudhari, Ashwin De Silva, Jingfeng Wu,
Carey E. Priebe
<!-- | Joshua T. Vogelstein <br> -->
<!-- [Microsoft Research](https://www.microsoft.com/en-us/research/): Weiwei Yang | Jonathan Larson | Bryan Tower | Chris White -->
<img src="images/neurodata_blue.png" width="20%" style="vertical-align: top">
<img src="images/jhu.png" width="8%" style="vertical-align: top">
---
<img src="images/L2M_hava.png" width="1000">
---
## Formalizing Learning
- Assume $(X\_i,Y\_i) \sim^{iid} P, \quad i \in [n]$
- Let $f((X_i,Y_i)^n) \rightarrow h_n$
- Let R (risk) be expected loss
- Let $\delta, \epsilon > 0$
- a learner .ye[weakly] learns if $\, \exists N$ s.t. $\forall n > N$
$$Pr[ R(h\_n) < R\_{chance} ] \geq 1 - \delta$$
- a learner .ye[strongly] learns if $\, \exists N$ s.t. $\forall n > N$
$$Pr[ R(h\_n) - R^* < \epsilon ] \geq 1 - \delta$$
- strong learner thm: if $f$ weakly learns, it also strongly learns
- implication: if your AI is doing ok, just get more data!
---
## So learning is solved?
- No.
- In lifelong learning, the samples are not all from the same distribution.
- Consider the simplest possible generalization:
- we have 2 datasets from two different distributions
- We want to know whether we learn about one from the other
- This is called '.ye[out of distribution]' learning
- Lifelong learning is a special (sequential) case
---
## Formalizing OOD Learning
- Assume $(X\_i,Y\_i) \sim^{iid} P, \quad i \in [n]$
- Assume $(X\_i,Y\_i) \sim^{iid} Q, \quad i \in ${$n+1, \ldots, n+m$}
- Let $f((X\_i,Y\_i)^{n+m}) \rightarrow h\_{n,m}$
- a learner .ye[weakly] learns about $P$ from $Q$ if $\, \exists M$ s.t. $\forall m > M$
$$Pr[ R(h\_{n,m}) < R(h\_n) ] \geq 1 - \delta$$
- a learner .ye[strongly] learns about $P$ from $Q$ if $\, \exists M$ s.t. $\forall m > M$
$$Pr[ R(h\_{n,m}) - R^* < \epsilon ] \geq 1 - \delta$$
- strong OOD learner thm: if $f$ weakly learns, it does not necessarily strongly learn
- implications: more data does not necessarily help!
.footnote[.small[Geisa, et al., Towards a theory of out-of-distribution learning. 2021]]
---
### A generalized definition of learning
- Let $S = (X\_i,Y\_i) \sim^{iid} P, \, \forall i \in [n]$
- Let $S' = (X\_i,Y\_i) \sim^{iid} Q, \, \forall i \in ${$n+1,\ldots, n+m$}
- Let $\mathcal{E}^t_f(S)$ := expected risk for task $t$ of learner $f$ using data $S$
- Let .ye[Learning Efficiency] be defined:
$$LE^t_f(S,S') = \frac{ \mathcal{E}^t_f(S)}{\mathcal{E}^t_f(S')}$$
- $f$ .ye[learns] about $P$ from $Q$ whenever $\log LE^t_f(S,S') >0$
- $f$ .ye[forgets] about $P$ from $Q$ whenever $\log LE^t_f(S,S') < 0$
- .ye[catastrophic forgetting]: $LE \ll 0$
- forward transfer, backward transfer, performance relative to STE, etc. are special cases
.footnote[.small[Vogelstein, et al. arxiv, 2020]]
---
### Fundamental Question of OOD Learning
- Assume $S = (X\_i,Y\_i) \sim^{iid} P, \quad i \in [n]$
- Assume $T = (X\_i,Y\_i) \sim^{iid} Q, \quad i \in ${$n+1, \ldots, n+m$}
- Assume $S \perp T$
<!-- the $n$ samples from $P$ and the $m$ samples from $Q$ are independent of each other -->
- What are the conditions on $P, Q, n, m$ such that $f$ can learn about $P$ from $Q$ (i.e., achieve $\log LE > 0$)?
We hope that for a given $P, Q, n$, there exists an $M$ such that if $m > M$, then data from $Q$ makes $f$ learn about (rather than forget about) $P$
---
## The simplest example
- $P$ is two gaussians
- $Q$ is two gaussians shifted by $\Delta$
<img src="https://github.com/neurodata/ood-tl/blob/main/reports/figures/gausstask_fig.png?raw=true" width="700">
<!-- ![:scale 100%](https://github.com/neurodata/ood-tl/blob/main/reports/figures/gausstask_fig.png?raw=true) -->
---
## A terrible thing
<img src="https://github.com/neurodata/ood-tl/blob/main/reports/figures/gaussian_task_analytical_plot.svg?raw=true" width="1000">
- Sometimes, for a fixed $\Delta$, error is non-monotonic wrt $m$
- implications: just because a little OOD data helps, does not mean that more will help!
.footnote[.small[de Silva, et al. Non-monotonic Out-of-Distribution Risk, in prep]]
---
## Not just in simulated data
- CIFAR 10
- Task 1: Bird vs. Cat
- Task 2: Deer vs. Dog
<img src="https://github.com/neurodata/ood-tl/blob/main/reports/figures/bridcat_deerdog.svg?raw=true" width="1000">
---
## Summary so far
- OOD learning is different in kind from in-distribution learning
- More data is not necessarily adequate to get arbitrary performance
- More data can actually hurt
- Time to think carefully about how much data of different kinds to use, and how to combine datasets, rather than throwing everything in a bucket
---
###
<img src="images/S1.png" width="1200">
---
###
<img src="images/S2.png" width="1200">
---
###
<img src="images/S3.png" width="1200">
---
.small[
### Publications
1. A. Geisa et al. [Towards a theory of out-of-distribution learning](https://arxiv.org/abs/2109.14501), arXiv, 2021.
1. J. T. Vogelstein et al. [Representation Ensembling for Synergistic Lifelong Learning with Quasilinear Complexity](https://arxiv.org/abs/2004.12908), arXiv, 2022.
1. H. Xu et al. [Simplest Streaming Trees](https://arxiv.org/abs/2110.08483), arXiv, 2022.
1. J. Dey et al. [Out-of-distribution and in-distribution posterior calibration using Kernel Density Polytopes](https://arxiv.org/abs/2201.13001), arXiv, 2022.
1. C. E. Priebe et al. [Modern Machine Learning: Partition and Vote](https://doi.org/10.1101/2020.04.29.068460), 2020.
1. R. Guo, et al. [Estimating Information-Theoretic Quantities with Uncertainty Forests](https://arxiv.org/abs/1907.00325). arXiv, 2019.
1. R. Perry, et al. [Manifold Forests: Closing the Gap on Neural Networks](https://openreview.net/forum?id=B1xewR4KvH). arXiv, 2019.
1. C. Shen and J. T. Vogelstein. [Decision Forests Induce Characteristic Kernels](https://arxiv.org/abs/1812.00029). arXiv, 2019.
1. M. Madhya, et al. [Geodesic Learning via Unsupervised Decision Forests](https://arxiv.org/abs/1907.02844). arXiv, 2019.
1. M. Madhya, et al. [PACSET (Packed Serialized Trees): Reducing Inference Latency for Tree Ensemble Deployment](https://arxiv.org/abs/2011.05383). arXiv, 2020.
<!-- ### Conferences -->
1. J. T. Vogelstein et al. A biological implementation of lifelong learning in the pursuit of artificial general intelligence. NAISys, 2020.
2. B. Pedigo et al. A quantitative comparison of a complete connectome to artificial intelligence architectures. NAISys, 2020.
3. L. Haoran, et al. [Lifelong Learning with Sketched Structural Regularization](https://arxiv.org/abs/2104.08604). ACML, 2021.
]
---
#### Acknowledgements
<!-- <div class="small-container">
<img src="faces/ebridge.jpg"/>
<div class="centered">Eric Bridgeford</div>
</div>
<div class="small-container">
<img src="faces/pedigo.jpg"/>
<div class="centered">Ben Pedigo</div>
</div>
<div class="small-container">
<img src="faces/jaewon.jpg"/>
<div class="centered">Jaewon Chung</div>
</div> -->
<div class="small-container">
<img src="faces/yummy.jpg"/>
<div class="centered">yummy</div>
</div>
<div class="small-container">
<img src="faces/lion.jpg"/>
<div class="centered">lion</div>
</div>
<div class="small-container">
<img src="faces/violet.jpg"/>
<div class="centered">baby girl</div>
</div>
<div class="small-container">
<img src="faces/family.jpg"/>
<div class="centered">family</div>
</div>
<div class="small-container">
<img src="faces/earth.jpg"/>
<div class="centered">earth</div>
</div>
<div class="small-container">
<img src="faces/milkyway.jpg"/>
<div class="centered">milkyway</div>
</div>
##### JHU
<div class="small-container">
<img src="faces/cep.png"/>
<div class="centered">Carey Priebe</div>
</div>
<div class="small-container">
<img src="faces/vova_lab/Vova.JPG"/>
<div class="centered">Vova Braverman</div>
</div>
<!-- <div class="small-container">
<img src="faces/randal.jpg"/>
<div class="centered">Randal Burns</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/cshen.jpg"/>
<div class="centered">Cencheng Shen</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/bruce_rosen.jpg"/>
<div class="centered">Bruce Rosen</div>
</div>
<div class="small-container">
<img src="faces/kent.jpg"/>
<div class="centered">Kent Kiehl</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/mim.jpg"/>
<div class="centered">Michael Miller</div>
</div>
<div class="small-container">
<img src="faces/dtward.jpg"/>
<div class="centered">Daniel Tward</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/vikram.jpg"/>
<div class="centered">Vikram Chandrashekhar</div>
</div>
<div class="small-container">
<img src="faces/drishti.jpg"/>
<div class="centered">Drishti Mannan</div>
</div> -->
<div class="small-container">
<img src="faces/jesse.jpg"/>
<div class="centered">Jesse Patsolic</div>
</div>
<!-- <div class="small-container">
<img src="faces/falk_ben.jpg"/>
<div class="centered">Benjamin Falk</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/kwame.jpg"/>
<div class="centered">Kwame Kutten</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/perlman.jpg"/>
<div class="centered">Eric Perlman</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/loftus.jpg"/>
<div class="centered">Alex Loftus</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/bcaffo.jpg"/>
<div class="centered">Brian Caffo</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/minh.jpg"/>
<div class="centered">Minh Tang</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/avanti.jpg"/>
<div class="centered">Avanti Athreya</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/vince.jpg"/>
<div class="centered">Vince Lyzinski</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/dpmcsuss.jpg"/>
<div class="centered">Daniel Sussman</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/youngser.jpg"/>
<div class="centered">Youngser Park</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/shangsi.jpg"/>
<div class="centered">Shangsi Wang</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/tyler.jpg"/>
<div class="centered">Tyler Tomita</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/james.jpg"/>
<div class="centered">James Brown</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/disa.jpg"/>
<div class="centered">Disa Mhembere</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/gkiar.jpg"/>
<div class="centered">Greg Kiar</div>
</div> -->
<!-- <div class="small-container">
<img src="faces/jeremias.png"/>
<div class="centered">Jeremias Sulam</div>
</div> -->
<div class="small-container">
<img src="faces/meghana.png"/>
<div class="centered">Meghana Madhya</div>
</div>
<!-- <div class="small-container">
<img src="faces/percy.png"/>
<div class="centered">Percy Li</div>
</div>
-->
<div class="small-container">
<img src="faces/hayden.png"/>
<div class="centered">Hayden Helm</div>
</div>
<div class="small-container">
<img src="faces/rguo.jpg"/>
<div class="centered">Richard Gou</div>
</div>
<div class="small-container">
<img src="faces/ronak.jpg"/>
<div class="centered">Ronak Mehta</div>
</div>
<div class="small-container">
<img src="faces/jayanta.jpg"/>
<div class="centered">Jayanta Dey</div>
</div>
<div class="small-container">
<img src="faces/will.jpg"/>
<div class="centered">Will LeVine</div>
</div>
<div class="small-container">
<img src="faces/hao.jpg"/>
<div class="centered">Haoyin Xu</div>
</div>
<div class="small-container">
<img src="faces/nhahn.jpg"/>
<div class="centered">Nick Hahn</div>
</div>
<div class="small-container">
<img src="faces/vova_lab/Haoran.JPG"/>
<div class="centered">Haoran Li</div>
</div>
<div class="small-container">
<img src="faces/vova_lab/Aditya.jpeg"/>
<div class="centered">Aditya Krishnan</div>
</div>
<div class="small-container">
<img src="faces/vova_lab/Jingfeng.jpg"/>
<div class="centered">Jingfeng Wu</div>
</div>
##### HRL & Microsoft Research
<div class="small-container">
<img src="faces/vova_lab/Soheil.PNG"/>
<div class="centered">Soheil Kolouri</div>
</div>
<div class="small-container">
<img src="faces/vova_lab/Praveen.JPG"/>
<div class="centered">Praveen K. Pilly</div>
</div>
<!-- ##### Microsoft Research -->
<div class="small-container">
<img src="faces/chwh-180x180.jpg"/>
<div class="centered">Chris White</div>
</div>
<div class="small-container">
<img src="faces/weiwei.jpg"/>
<div class="centered">Weiwei Yang</div>
</div>
<div class="small-container">
<img src="faces/jolarso150px.png"/>
<div class="centered">Jonathan Larson</div>
</div>
<div class="small-container">
<img src="faces/brtower-180x180.jpg"/>
<div class="centered">Bryan Tower</div>
</div>
##### DARPA L2M
<!-- Hava, Ben, Robert, Jennifer, Ted. -->
<!-- {[BME](https://www.bme.jhu.edu/),[CIS](http://cis.jhu.edu/), [ICM](https://icm.jhu.edu/), [KNDI](http://kavlijhu.org/)}@[JHU](https://www.jhu.edu/) | [neurodata](https://neurodata.io) -->
<!-- <br> -->
<!-- [jovo@jhu.edu](mailto:[email protected]) | <http://neurodata.io/talks> | [@neuro_data](https://twitter.com/neuro_data) -->
</div>
<!-- <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;"/> -->
---
background-image: url(images/l_and_v.jpeg)
.footnote[Questions?]
---
class: middle
# .center[Appendix]
---
### Biological learning is on top
<img src="images/learning-table.png" width="1000">
---
### Spoken Digit dataset
.pull-left[
- *Spoken Digit* contains recording from 6 different speakers.
- Each digit has 50 recordings (3000 total recordings).
- For each recording spectrogram was extracted using using Hanning windows of duration 16 ms with an overlap of 4 ms.
- The spectrograms were resized down to 28×28.
]
.pull-right[
<img src="images/spectrogram.png" style="left:500px; width:400px;"/>
]
---
### Synergistic Algorithms on Spoken Digit Task
<img src="images/spoken_digit.png" width="1000">
<!-- ![:scale 105%]() -->
</textarea>
<!-- <script src="https://gnab.github.io/remark/downloads/remark-latest.min.js"></script> -->
<!-- <script src="remark-latest.min.js"></script> -->
<script src="remark-latest.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/KaTeX/0.5.1/katex.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/KaTeX/0.5.1/contrib/auto-render.min.js"></script>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/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 },
]
});
}
// remark.macros.scale = function (percentage) {
// var url = this;
// return '<img src="' + url + '" style="width: ' + percentage + '" />';
// };
// var slideshow = remark.create({
// Set the slideshow display ratio
// Default: '4:3'
// Alternatives: '16:9', ...
// {
// ratio: '16:9',
// });
var slideshow = remark.create(options, renderMath);
</script>
</body>
</html>