File size: 43,617 Bytes
323a418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
WEBVTT

00:00.000 --> 00:04.320
 What difference between biological neural networks and artificial neural networks

00:04.320 --> 00:07.680
 is most mysterious, captivating and profound for you?

00:11.120 --> 00:15.280
 First of all, there's so much we don't know about biological neural networks,

00:15.280 --> 00:21.840
 and that's very mysterious and captivating because maybe it holds the key to improving

00:21.840 --> 00:29.840
 artificial neural networks. One of the things I studied recently is something that

00:29.840 --> 00:36.160
 we don't know how biological neural networks do, but would be really useful for artificial ones,

00:37.120 --> 00:43.440
 is the ability to do credit assignment through very long time spans.

00:44.080 --> 00:49.680
 There are things that we can in principle do with artificial neural nets, but it's not very

00:49.680 --> 00:55.920
 convenient and it's not biologically plausible. And this mismatch, I think this kind of mismatch,

00:55.920 --> 01:03.600
 maybe an interesting thing to study, to A, understand better how brains might do these

01:03.600 --> 01:08.720
 things because we don't have good corresponding theories with artificial neural nets, and B,

01:10.240 --> 01:19.040
 maybe provide new ideas that we could explore about things that brain do differently and

01:19.040 --> 01:22.160
 that we could incorporate in artificial neural nets.

01:22.160 --> 01:27.680
 So let's break credit assignment up a little bit. So what? It's a beautifully technical term,

01:27.680 --> 01:34.560
 but it could incorporate so many things. So is it more on the RNN memory side,

01:35.840 --> 01:39.760
 thinking like that, or is it something about knowledge, building up common sense knowledge

01:39.760 --> 01:46.560
 over time, or is it more in the reinforcement learning sense that you're picking up rewards

01:46.560 --> 01:50.080
 over time for a particular to achieve a certain kind of goal?

01:50.080 --> 01:58.080
 So I was thinking more about the first two meanings whereby we store all kinds of memories,

01:59.120 --> 02:09.680
 episodic memories in our brain, which we can access later in order to help us both infer

02:10.560 --> 02:19.520
 causes of things that we are observing now and assign credit to decisions or interpretations

02:19.520 --> 02:26.960
 we came up with a while ago when those memories were stored. And then we can change the way we

02:26.960 --> 02:34.800
 would have reacted or interpreted things in the past, and now that's credit assignment used for learning.

02:36.320 --> 02:43.760
 So in which way do you think artificial neural networks, the current LSTM,

02:43.760 --> 02:52.240
 the current architectures are not able to capture the presumably you're thinking of very long term?

02:52.240 --> 03:00.720
 Yes. So current, the current nets are doing a fairly good jobs for sequences with dozens or say

03:00.720 --> 03:06.560
 hundreds of time steps. And then it gets sort of harder and harder and depending on what you

03:06.560 --> 03:13.120
 have to remember and so on as you consider longer durations. Whereas humans seem to be able to

03:13.120 --> 03:18.080
 do credit assignment through essentially arbitrary times like I could remember something I did last

03:18.080 --> 03:23.360
 year. And then now because I see some new evidence, I'm going to change my mind about

03:23.360 --> 03:29.040
 the way I was thinking last year, and hopefully not do the same mistake again.

03:31.040 --> 03:36.800
 I think a big part of that is probably forgetting. You're only remembering the really important

03:36.800 --> 03:43.680
 things that's very efficient forgetting. Yes. So there's a selection of what we remember.

03:43.680 --> 03:49.120
 And I think there are really cool connection to higher level cognitions here regarding

03:49.120 --> 03:55.760
 consciousness, deciding and emotions. So deciding what comes to consciousness and what gets stored

03:55.760 --> 04:04.800
 in memory, which are not trivial either. So you've been at the forefront there all along

04:04.800 --> 04:10.800
 showing some of the amazing things that neural networks, deep neural networks can do in the

04:10.800 --> 04:16.560
 field of artificial intelligence is just broadly in all kinds of applications. But we can talk

04:16.560 --> 04:23.200
 about that forever. But what in your view, because we're thinking towards the future is the weakest

04:23.200 --> 04:29.120
 aspect of the way deep neural networks represent the world. What is that? What is in your view

04:29.120 --> 04:41.200
 is missing? So current state of the art neural nets trained on large quantities of images or texts

04:43.840 --> 04:49.760
 have some level of understanding of what explains those data sets, but it's very

04:49.760 --> 05:01.440
 basic. It's very low level. And it's not nearly as robust and abstract and general as our understanding.

05:02.960 --> 05:09.760
 Okay, so that doesn't tell us how to fix things. But I think it encourages us to think about

05:09.760 --> 05:21.200
 how we can maybe train our neural nets differently, so that they would focus, for example, on causal

05:21.200 --> 05:30.000
 explanations, something that we don't do currently with neural net training. Also, one thing I'll

05:30.000 --> 05:37.920
 talk about in my talk this afternoon is instead of learning separately from images and videos on

05:37.920 --> 05:45.600
 one hand and from texts on the other hand, we need to do a better job of jointly learning about

05:45.600 --> 05:54.320
 language and about the world to which it refers. So that, you know, both sides can help each other.

05:54.880 --> 06:02.480
 We need to have good world models in our neural nets for them to really understand sentences

06:02.480 --> 06:10.000
 which talk about what's going on in the world. And I think we need language input to help

06:10.640 --> 06:17.760
 provide clues about what high level concepts like semantic concepts should be represented

06:17.760 --> 06:26.400
 at the top levels of these neural nets. In fact, there is evidence that the purely unsupervised

06:26.400 --> 06:33.840
 learning of representations doesn't give rise to high level representations that are as powerful

06:33.840 --> 06:40.320
 as the ones we're getting from supervised learning. And so the clues we're getting just with the labels,

06:40.320 --> 06:46.960
 not even sentences, is already very powerful. Do you think that's an architecture challenge

06:46.960 --> 06:55.920
 or is it a data set challenge? Neither. I'm tempted to just end it there.

07:02.960 --> 07:06.800
 Of course, data sets and architectures are something you want to always play with. But

07:06.800 --> 07:13.040
 I think the crucial thing is more the training objectives, the training frameworks. For example,

07:13.040 --> 07:20.240
 going from passive observation of data to more active agents, which

07:22.320 --> 07:27.280
 learn by intervening in the world, the relationships between causes and effects,

07:28.480 --> 07:36.240
 the sort of objective functions which could be important to allow the highest level

07:36.240 --> 07:44.000
 of explanations to rise from the learning, which I don't think we have now. The kinds of

07:44.000 --> 07:50.320
 objective functions which could be used to reward exploration, the right kind of exploration. So

07:50.320 --> 07:56.160
 these kinds of questions are neither in the data set nor in the architecture, but more in

07:56.800 --> 08:03.920
 how we learn under what objectives and so on. Yeah, that's a, I've heard you mention in several

08:03.920 --> 08:08.080
 contexts, the idea of sort of the way children learn, they interact with objects in the world.

08:08.080 --> 08:15.040
 And it seems fascinating because in some sense, except with some cases in reinforcement learning,

08:15.760 --> 08:23.600
 that idea is not part of the learning process in artificial neural networks. It's almost like

08:24.320 --> 08:33.120
 do you envision something like an objective function saying, you know what, if you poke this

08:33.120 --> 08:38.800
 object in this kind of way, it would be really helpful for me to further, further learn.

08:39.920 --> 08:44.880
 Sort of almost guiding some aspect of learning. Right, right, right. So I was talking to Rebecca

08:44.880 --> 08:54.240
 Sachs just an hour ago and she was talking about lots and lots of evidence from infants seem to

08:54.240 --> 09:04.880
 clearly pick what interests them in a directed way. And so they're not passive learners.

09:04.880 --> 09:11.680
 They, they focus their attention on aspects of the world, which are most interesting,

09:11.680 --> 09:17.760
 surprising in a non trivial way that makes them change their theories of the world.

09:17.760 --> 09:29.120
 So that's a fascinating view of the future progress. But on a more maybe boring question,

09:30.000 --> 09:37.440
 do you think going deeper and larger? So do you think just increasing the size of the things

09:37.440 --> 09:43.520
 that have been increasing a lot in the past few years will, will also make significant progress?

09:43.520 --> 09:49.760
 So some of the representational issues that you, you mentioned, they're kind of shallow

09:50.560 --> 09:54.880
 in some sense. Oh, you mean in the sense of abstraction,

09:54.880 --> 09:59.040
 abstract in the sense of abstraction, they're not getting some, I don't think that having

10:00.400 --> 10:05.520
 more depth in the network in the sense of instead of 100 layers, we have 10,000 is going to solve

10:05.520 --> 10:13.120
 our problem. You don't think so? Is that obvious to you? Yes. What is clear to me is that

10:13.120 --> 10:21.600
 engineers and companies and labs, grad students will continue to tune architectures and explore

10:21.600 --> 10:27.520
 all kinds of tweaks to make the current state of the art slightly ever slightly better. But

10:27.520 --> 10:31.840
 I don't think that's going to be nearly enough. I think we need some fairly drastic changes in

10:31.840 --> 10:39.680
 the way that we're considering learning to achieve the goal that these learners actually

10:39.680 --> 10:45.680
 understand in a deep way the environment in which they are, you know, observing and acting.

10:46.480 --> 10:51.920
 But I guess I was trying to ask a question that's more interesting than just more layers

10:53.040 --> 11:00.800
 is basically once you figure out a way to learn through interacting, how many parameters does

11:00.800 --> 11:07.760
 it take to store that information? So I think our brain is quite bigger than most neural networks.

11:07.760 --> 11:13.120
 Right, right. Oh, I see what you mean. Oh, I'm with you there. So I agree that in order to

11:14.240 --> 11:19.760
 build neural nets with the kind of broad knowledge of the world that typical adult humans have,

11:20.960 --> 11:24.880
 probably the kind of computing power we have now is going to be insufficient.

11:25.600 --> 11:30.320
 So the good news is there are hardware companies building neural net chips. And so

11:30.320 --> 11:39.280
 it's going to get better. However, the good news in a way, which is also a bad news, is that even

11:39.280 --> 11:47.840
 our state of the art deep learning methods fail to learn models that understand even very simple

11:47.840 --> 11:53.680
 environments like some grid worlds that we have built. Even these fairly simple environments,

11:53.680 --> 11:57.120
 I mean, of course, if you train them with enough examples, eventually they get it,

11:57.120 --> 12:05.200
 but it's just like instead of what humans might need just dozens of examples, these things will

12:05.200 --> 12:12.720
 need millions, right, for very, very, very simple tasks. And so I think there's an opportunity

12:13.520 --> 12:18.080
 for academics who don't have the kind of computing power that say Google has

12:19.280 --> 12:25.360
 to do really important and exciting research to advance the state of the art in training

12:25.360 --> 12:32.720
 frameworks, learning models, agent learning in even simple environments that are synthetic,

12:33.440 --> 12:37.200
 that seem trivial, but yet current machine learning fails on.

12:38.240 --> 12:48.240
 We talked about priors and common sense knowledge. It seems like we humans take a lot of knowledge

12:48.240 --> 12:57.040
 for granted. So what's your view of these priors of forming this broad view of the world, this

12:57.040 --> 13:02.560
 accumulation of information, and how we can teach neural networks or learning systems to pick that

13:02.560 --> 13:10.880
 knowledge up? So knowledge, you know, for a while, the artificial intelligence, maybe in the 80,

13:10.880 --> 13:16.880
 like there's a time where knowledge representation, knowledge, acquisition, expert systems, I mean,

13:16.880 --> 13:24.080
 though, the symbolic AI was a view, was an interesting problem set to solve. And it was kind

13:24.080 --> 13:29.440
 of put on hold a little bit, it seems like because it doesn't work. It doesn't work. That's right.

13:29.440 --> 13:37.840
 But that's right. But the goals of that remain important. Yes, remain important. And how do you

13:37.840 --> 13:45.920
 think those goals can be addressed? Right. So first of all, I believe that one reason why the

13:45.920 --> 13:52.560
 classical expert systems approach failed is because a lot of the knowledge we have, so you talked

13:52.560 --> 14:01.760
 about common sense and tuition, there's a lot of knowledge like this, which is not consciously

14:01.760 --> 14:06.320
 accessible. There are lots of decisions we're taking that we can't really explain, even if

14:06.320 --> 14:16.160
 sometimes we make up a story. And that knowledge is also necessary for machines to take good

14:16.160 --> 14:22.320
 decisions. And that knowledge is hard to codify in expert systems, rule based systems, and, you

14:22.320 --> 14:27.920
 know, classical AI formalism. And there are other issues, of course, with the old AI, like,

14:29.680 --> 14:34.320
 not really good ways of handling uncertainty, I would say something more subtle,

14:34.320 --> 14:40.480
 which we understand better now, but I think still isn't enough in the minds of people.

14:41.360 --> 14:48.480
 There's something really powerful that comes from distributed representations, the thing that really

14:49.120 --> 14:58.480
 makes neural nets work so well. And it's hard to replicate that kind of power in a symbolic world.

14:58.480 --> 15:05.200
 The knowledge in expert systems and so on is nicely decomposed into like a bunch of rules.

15:05.760 --> 15:11.280
 Whereas if you think about a neural net, it's the opposite. You have this big blob of parameters

15:11.280 --> 15:16.480
 which work intensely together to represent everything the network knows. And it's not

15:16.480 --> 15:22.880
 sufficiently factorized. And so I think this is one of the weaknesses of current neural nets,

15:22.880 --> 15:30.080
 that we have to take lessons from classical AI in order to bring in another kind of

15:30.080 --> 15:35.920
 compositionality, which is common in language, for example, and in these rules. But that isn't

15:35.920 --> 15:45.040
 so native to neural nets. And on that line of thinking, disentangled representations. Yes. So

15:46.320 --> 15:51.680
 let me connect with disentangled representations. If you might, if you don't mind. Yes, exactly.

15:51.680 --> 15:58.080
 Yeah. So for many years, I thought, and I still believe that it's really important that we come

15:58.080 --> 16:04.080
 up with learning algorithms, either unsupervised or supervised, but reinforcement, whatever,

16:04.720 --> 16:11.600
 that build representations in which the important factors, hopefully causal factors are nicely

16:11.600 --> 16:16.240
 separated and easy to pick up from the representation. So that's the idea of disentangled

16:16.240 --> 16:22.560
 representations. It says transfer the data into a space where everything becomes easy, we can maybe

16:22.560 --> 16:29.360
 just learn with linear models about the things we care about. And I still think this is important,

16:29.360 --> 16:36.880
 but I think this is missing out on a very important ingredient, which classical AI systems can remind

16:36.880 --> 16:41.920
 us of. So let's say we have these disentangled representations, you still need to learn about

16:41.920 --> 16:47.120
 the, the relationships between the variables, those high level semantic variables, they're not

16:47.120 --> 16:52.000
 going to be independent. I mean, this is like too much of an assumption. They're going to have some

16:52.000 --> 16:56.400
 interesting relationships that allow to predict things in the future to explain what happened in

16:56.400 --> 17:01.840
 the past. The kind of knowledge about those relationships in a classical AI system is

17:01.840 --> 17:06.640
 encoded in the rules, like a rule is just like a little piece of knowledge that says, oh, I have

17:06.640 --> 17:12.160
 these two, three, four variables that are linked in this interesting way. Then I can say something

17:12.160 --> 17:17.280
 about one or two of them given a couple of others, right? In addition to disentangling the,

17:18.880 --> 17:23.520
 the elements of the representation, which are like the variables in a rule based system,

17:24.080 --> 17:33.200
 you also need to disentangle the, the mechanisms that relate those variables to each other.

17:33.200 --> 17:37.760
 So like the rules. So if the rules are neatly separated, like each rule is, you know, living

17:37.760 --> 17:44.960
 on its own. And when I, I change a rule because I'm learning, it doesn't need to break other rules.

17:44.960 --> 17:49.280
 Whereas current neural nets, for example, are very sensitive to what's called catastrophic

17:49.280 --> 17:54.800
 forgetting, where after I've learned some things, and then they learn new things, they can destroy

17:54.800 --> 18:00.480
 the old things that I had learned, right? If the knowledge was better factorized and, and

18:00.480 --> 18:08.240
 and separated disentangled, then you would avoid a lot of that. Now you can't do this in the

18:08.880 --> 18:17.200
 sensory domain, but my idea in like a pixel space, but, but my idea is that when you project the

18:17.200 --> 18:22.560
 data in the right semantic space, it becomes possible to now represent this extra knowledge

18:23.440 --> 18:27.760
 beyond the transformation from input to representations, which is how representations

18:27.760 --> 18:33.120
 act on each other and predict the future and so on, in a way that can be neatly

18:34.560 --> 18:38.560
 disentangled. So now it's the rules that are disentangled from each other and not just the

18:38.560 --> 18:43.680
 variables that are disentangled from each other. And you draw distinction between semantic space

18:43.680 --> 18:48.400
 and pixel, like, does there need to be an architectural difference? Well, yeah. So, so

18:48.400 --> 18:51.840
 there's the sensory space like pixels, which where everything is entangled,

18:51.840 --> 18:58.000
 and the information, like the variables are completely interdependent in very complicated

18:58.000 --> 19:03.760
 ways. And also computation, like the, it's not just variables, it's also how they are

19:03.760 --> 19:10.240
 related to each other is, is all intertwined. But, but I'm hypothesizing that in the right

19:10.240 --> 19:16.800
 high level representation space, both the variables and how they relate to each other

19:16.800 --> 19:22.960
 can be disentangled and that will provide a lot of generalization power. Generalization power.

19:22.960 --> 19:29.760
 Yes. Distribution of the test set, it's assumed to be the same as a distribution of the training

19:29.760 --> 19:36.640
 set. Right. This is where current machine learning is too weak. It doesn't tell us anything,

19:36.640 --> 19:41.120
 is not able to tell us anything about how our neural nets, say, are going to generalize to a

19:41.120 --> 19:46.160
 new distribution. And, and, you know, people may think, well, but there's nothing we can say if

19:46.160 --> 19:51.840
 we don't know what the new distribution will be. The truth is, humans are able to generalize to

19:51.840 --> 19:56.560
 new distributions. Yeah, how are we able to do that? So yeah, because there is something, these

19:56.560 --> 20:00.720
 new distributions, even though they could look very different from the training distributions,

20:01.520 --> 20:05.360
 they have things in common. So let me give you a concrete example. You read a science fiction

20:05.360 --> 20:12.560
 novel, the science fiction novel, maybe, you know, brings you in some other planet where

20:12.560 --> 20:17.760
 things look very different on the surface, but it's still the same laws of physics.

20:18.560 --> 20:21.440
 All right. And so you can read the book and you understand what's going on.

20:22.960 --> 20:29.200
 So the distribution is very different. But because you can transport a lot of the knowledge you had

20:29.200 --> 20:35.680
 from Earth about the underlying cause and effect relationships and physical mechanisms and all

20:35.680 --> 20:40.880
 that, and maybe even social interactions, you can now make sense of what is going on on this

20:40.880 --> 20:43.920
 planet where like visually, for example, things are totally different.

20:45.920 --> 20:52.000
 Taking that analogy further and distorting it, let's enter a science fiction world of, say,

20:52.000 --> 21:00.720
 Space Odyssey 2001 with Hal. Yeah. Or maybe, which is probably one of my favorite AI movies.

21:00.720 --> 21:06.080
 Me too. And then there's another one that a lot of people love that may be a little bit outside

21:06.080 --> 21:13.120
 of the AI community is Ex Machina. I don't know if you've seen it. Yes. By the way, what are your

21:13.120 --> 21:19.600
 reviews on that movie? Are you able to enjoy it? So there are things I like and things I hate.

21:21.120 --> 21:25.760
 So let me, you could talk about that in the context of a question I want to ask,

21:25.760 --> 21:31.920
 which is there's quite a large community of people from different backgrounds off and outside of AI

21:31.920 --> 21:36.480
 who are concerned about existential threat of artificial intelligence. Right. You've seen

21:36.480 --> 21:41.920
 now this community develop over time. You've seen you have a perspective. So what do you think is

21:41.920 --> 21:47.680
 the best way to talk about AI safety, to think about it, to have discourse about it within AI

21:47.680 --> 21:53.920
 community and outside and grounded in the fact that Ex Machina is one of the main sources of

21:53.920 --> 21:59.040
 information for the general public about AI. So I think you're putting it right. There's a big

21:59.040 --> 22:04.400
 difference between the sort of discussion we ought to have within the AI community

22:05.200 --> 22:11.600
 and the sort of discussion that really matter in the general public. So I think the picture of

22:11.600 --> 22:19.040
 Terminator and, you know, AI loose and killing people and super intelligence that's going to

22:19.040 --> 22:26.320
 destroy us, whatever we try, isn't really so useful for the public discussion because

22:26.320 --> 22:32.960
 for the public discussion that things I believe really matter are the short term and

22:32.960 --> 22:40.560
 mini term, very likely negative impacts of AI on society, whether it's from security,

22:40.560 --> 22:45.680
 like, you know, big brother scenarios with face recognition or killer robots, or the impact on

22:45.680 --> 22:52.400
 the job market, or concentration of power and discrimination, all kinds of social issues,

22:52.400 --> 22:58.240
 which could actually, some of them could really threaten democracy, for example.

22:58.800 --> 23:04.000
 Just to clarify, when you said killer robots, you mean autonomous weapons as a weapon system?

23:04.000 --> 23:10.400
 Yes, I don't mean, no, that's right. So I think these short and medium term concerns

23:11.280 --> 23:18.560
 should be important parts of the public debate. Now, existential risk, for me, is a very unlikely

23:18.560 --> 23:26.880
 consideration, but still worth academic investigation. In the same way that you could say,

23:26.880 --> 23:32.640
 should we study what could happen if meteorite, you know, came to earth and destroyed it.

23:32.640 --> 23:37.680
 So I think it's very unlikely that this is going to happen in or happen in a reasonable future.

23:37.680 --> 23:45.520
 It's very, the sort of scenario of an AI getting loose goes against my understanding of at least

23:45.520 --> 23:50.160
 current machine learning and current neural nets and so on. It's not plausible to me.

23:50.160 --> 23:54.320
 But of course, I don't have a crystal ball and who knows what AI will be in 50 years from now.

23:54.320 --> 23:59.280
 So I think it is worth that scientists study those problems. It's just not a pressing question,

23:59.280 --> 24:04.880
 as far as I'm concerned. So before I continue down that line, I have a few questions there, but

24:06.640 --> 24:11.440
 what do you like and not like about X Machina as a movie? Because I actually watched it for the

24:11.440 --> 24:17.840
 second time and enjoyed it. I hated it the first time and I enjoyed it quite a bit more the second

24:17.840 --> 24:26.080
 time when I sort of learned to accept certain pieces of it. See it as a concept movie. What

24:26.080 --> 24:36.160
 was your experience? What were your thoughts? So the negative is the picture it paints of science

24:36.160 --> 24:41.760
 is totally wrong. Science in general and AI in particular. Science is not happening

24:43.120 --> 24:51.840
 in some hidden place by some really smart guy. One person. One person. This is totally unrealistic.

24:51.840 --> 24:58.240
 This is not how it happens. Even a team of people in some isolated place will not make it.

24:58.240 --> 25:07.920
 Science moves by small steps thanks to the collaboration and community of a large number

25:07.920 --> 25:16.000
 of people interacting and all the scientists who are expert in their field kind of know what is

25:16.000 --> 25:24.000
 going on even in the industrial labs. Information flows and leaks and so on. And the spirit of

25:24.000 --> 25:30.320
 it is very different from the way science is painted in this movie. Yeah, let me ask on that

25:30.320 --> 25:36.400
 point. It's been the case to this point that kind of even if the research happens inside

25:36.400 --> 25:42.000
 Google or Facebook, inside companies, it still kind of comes out. Do you think that will always be

25:42.000 --> 25:48.960
 the case with AI? Is it possible to bottle ideas to the point where there's a set of breakthroughs

25:48.960 --> 25:53.120
 that go completely undiscovered by the general research community? Do you think that's even

25:53.120 --> 26:02.240
 possible? It's possible, but it's unlikely. It's not how it is done now. It's not how I can force

26:02.240 --> 26:13.120
 it in in the foreseeable future. But of course, I don't have a crystal ball. And so who knows,

26:13.120 --> 26:18.240
 this is science fiction after all. But but usually ominous that the lights went off during

26:18.240 --> 26:24.320
 during that discussion. So the problem again, there's a you know, one thing is the movie and

26:24.320 --> 26:28.720
 you could imagine all kinds of science fiction. The problem with for me, maybe similar to the

26:28.720 --> 26:37.120
 question about existential risk is that this kind of movie paints such a wrong picture of what is

26:37.120 --> 26:43.520
 actual, you know, the actual science and how it's going on that that it can have unfortunate effects

26:43.520 --> 26:49.040
 on people's understanding of current science. And so that's kind of sad.

26:50.560 --> 26:56.800
 There's an important principle in research, which is diversity. So in other words,

26:58.000 --> 27:02.720
 research is exploration, research is exploration in the space of ideas. And different people

27:03.440 --> 27:09.920
 will focus on different directions. And this is not just good, it's essential. So I'm totally fine

27:09.920 --> 27:16.640
 with people exploring directions that are contrary to mine or look orthogonal to mine.

27:18.560 --> 27:24.880
 I am more than fine, I think it's important. I and my friends don't claim we have universal

27:24.880 --> 27:29.680
 truth about what will especially about what will happen in the future. Now that being said,

27:30.320 --> 27:37.600
 we have our intuitions and then we act accordingly, according to where we think we can be most useful

27:37.600 --> 27:43.360
 and where society has the most to gain or to lose. We should have those debates and

27:45.920 --> 27:50.080
 and not end up in a society where there's only one voice and one way of thinking and

27:51.360 --> 27:59.120
 research money is spread out. So this agreement is a sign of good research, good science. So

27:59.120 --> 28:08.560
 yes. The idea of bias in the human sense of bias. How do you think about instilling in machine

28:08.560 --> 28:15.440
 learning something that's aligned with human values in terms of bias? We intuitively assume

28:15.440 --> 28:21.680
 beings have a concept of what bias means, of what fundamental respect for other human beings means,

28:21.680 --> 28:25.280
 but how do we instill that into machine learning systems, do you think?

28:25.280 --> 28:32.720
 So I think there are short term things that are already happening and then there are long term

28:32.720 --> 28:39.040
 things that we need to do. In the short term, there are techniques that have been proposed and

28:39.040 --> 28:44.800
 I think will continue to be improved and maybe alternatives will come up to take data sets

28:45.600 --> 28:51.200
 in which we know there is bias, we can measure it. Pretty much any data set where humans are

28:51.200 --> 28:56.080
 being observed taking decisions will have some sort of bias discrimination against particular

28:56.080 --> 29:04.000
 groups and so on. And we can use machine learning techniques to try to build predictors, classifiers

29:04.000 --> 29:11.920
 that are going to be less biased. We can do it for example using adversarial methods to make our

29:11.920 --> 29:19.520
 systems less sensitive to these variables we should not be sensitive to. So these are clear,

29:19.520 --> 29:24.240
 well defined ways of trying to address the problem, maybe they have weaknesses and more

29:24.240 --> 29:30.400
 research is needed and so on, but I think in fact they're sufficiently mature that governments should

29:30.400 --> 29:36.160
 start regulating companies where it matters say like insurance companies so that they use those

29:36.160 --> 29:43.840
 techniques because those techniques will probably reduce the bias, but at a cost for example maybe

29:43.840 --> 29:47.920
 their predictions will be less accurate and so companies will not do it until you force them.

29:47.920 --> 29:56.000
 All right, so this is short term. Long term, I'm really interested in thinking how we can

29:56.000 --> 30:02.160
 instill moral values into computers. Obviously this is not something we'll achieve in the next five

30:02.160 --> 30:11.680
 or 10 years. There's already work in detecting emotions for example in images and sounds and

30:11.680 --> 30:21.520
 texts and also studying how different agents interacting in different ways may correspond to

30:22.960 --> 30:30.000
 patterns of say injustice which could trigger anger. So these are things we can do in the

30:30.000 --> 30:42.160
 medium term and eventually train computers to model for example how humans react emotionally. I would

30:42.160 --> 30:49.920
 say the simplest thing is unfair situations which trigger anger. This is one of the most basic

30:49.920 --> 30:55.360
 emotions that we share with other animals. I think it's quite feasible within the next few years so

30:55.360 --> 31:00.800
 we can build systems that can detect these kind of things to the extent unfortunately that they

31:00.800 --> 31:07.840
 understand enough about the world around us which is a long time away but maybe we can initially do

31:07.840 --> 31:14.800
 this in virtual environments so you can imagine like a video game where agents interact in some

31:14.800 --> 31:21.760
 ways and then some situations trigger an emotion. I think we could train machines to detect those

31:21.760 --> 31:27.920
 situations and predict that the particular emotion will likely be felt if a human was playing one

31:27.920 --> 31:34.080
 of the characters. You have shown excitement and done a lot of excellent work with unsupervised

31:34.080 --> 31:42.800
 learning but there's been a lot of success on the supervised learning. One of the things I'm

31:42.800 --> 31:48.800
 really passionate about is how humans and robots work together and in the context of supervised

31:48.800 --> 31:54.800
 learning that means the process of annotation. Do you think about the problem of annotation of

31:55.520 --> 32:04.080
 put in a more interesting way is humans teaching machines? Yes, I think it's an important subject.

32:04.880 --> 32:11.280
 Reducing it to annotation may be useful for somebody building a system tomorrow but

32:12.560 --> 32:17.600
 longer term the process of teaching I think is something that deserves a lot more attention

32:17.600 --> 32:21.840
 from the machine learning community so there are people of coin the term machine teaching.

32:22.560 --> 32:30.480
 So what are good strategies for teaching a learning agent and can we design, train a system

32:30.480 --> 32:38.000
 that is going to be a good teacher? So in my group we have a project called a BBI or BBI game

32:38.640 --> 32:46.000
 where there is a game or a scenario where there's a learning agent and a teaching agent

32:46.000 --> 32:54.400
 presumably the teaching agent would eventually be a human but we're not there yet and the

32:56.000 --> 33:00.880
 role of the teacher is to use its knowledge of the environment which it can acquire using

33:00.880 --> 33:09.680
 whatever way brute force to help the learner learn as quickly as possible. So the learner

33:09.680 --> 33:13.920
 is going to try to learn by itself maybe using some exploration and whatever

33:13.920 --> 33:21.520
 but the teacher can choose, can have an influence on the interaction with the learner

33:21.520 --> 33:28.960
 so as to guide the learner maybe teach it the things that the learner has most trouble with

33:28.960 --> 33:34.320
 or just add the boundary between what it knows and doesn't know and so on. So there's a tradition

33:34.320 --> 33:41.280
 of these kind of ideas from other fields and like tutorial systems for example and AI

33:41.280 --> 33:46.880
 and of course people in the humanities have been thinking about these questions but I think

33:46.880 --> 33:52.560
 it's time that machine learning people look at this because in the future we'll have more and more

33:53.760 --> 33:59.680
 human machine interaction with the human in the loop and I think understanding how to make this

33:59.680 --> 34:04.080
 work better. Oh the problems around that are very interesting and not sufficiently addressed.

34:04.080 --> 34:11.440
 You've done a lot of work with language too, what aspect of the traditionally formulated

34:11.440 --> 34:17.040
 touring test, a test of natural language understanding in generation in your eyes is the

34:17.040 --> 34:22.960
 most difficult of conversation, what in your eyes is the hardest part of conversation to solve for

34:22.960 --> 34:30.640
 machines. So I would say it's everything having to do with the non linguistic knowledge which

34:30.640 --> 34:36.400
 implicitly you need in order to make sense of sentences. Things like the winner grad schemas

34:36.400 --> 34:42.400
 so these sentences that are semantically ambiguous. In other words you need to understand enough about

34:42.400 --> 34:48.720
 the world in order to really interpret properly those sentences. I think these are interesting

34:48.720 --> 34:55.840
 challenges for machine learning because they point in the direction of building systems that

34:55.840 --> 35:02.880
 both understand how the world works and there's causal relationships in the world and associate

35:03.520 --> 35:09.760
 that knowledge with how to express it in language either for reading or writing.

35:11.840 --> 35:17.600
 You speak French? Yes, it's my mother tongue. It's one of the romance languages. Do you think

35:17.600 --> 35:23.040
 passing the touring test and all the underlying challenges we just mentioned depend on language?

35:23.040 --> 35:28.000
 Do you think it might be easier in French than it is in English or is independent of language?

35:28.800 --> 35:37.680
 I think it's independent of language. I would like to build systems that can use the same

35:37.680 --> 35:45.840
 principles, the same learning mechanisms to learn from human agents, whatever their language.

35:45.840 --> 35:53.600
 Well, certainly us humans can talk more beautifully and smoothly in poetry. So I'm Russian originally.

35:53.600 --> 36:01.360
 I know poetry in Russian is maybe easier to convey complex ideas than it is in English

36:02.320 --> 36:09.520
 but maybe I'm showing my bias and some people could say that about French. But of course the

36:09.520 --> 36:16.400
 goal ultimately is our human brain is able to utilize any kind of those languages to use them

36:16.400 --> 36:21.040
 as tools to convey meaning. Yeah, of course there are differences between languages and maybe some

36:21.040 --> 36:25.920
 are slightly better at some things but in the grand scheme of things where we're trying to understand

36:25.920 --> 36:31.040
 how the brain works and language and so on, I think these differences are minute.

36:31.040 --> 36:42.880
 So you've lived perhaps through an AI winter of sorts. Yes. How did you stay warm and continue

36:42.880 --> 36:48.480
 with your research? Stay warm with friends. With friends. Okay, so it's important to have friends

36:48.480 --> 36:57.200
 and what have you learned from the experience? Listen to your inner voice. Don't, you know, be

36:57.200 --> 37:07.680
 trying to just please the crowds and the fashion and if you have a strong intuition about something

37:08.480 --> 37:15.520
 that is not contradicted by actual evidence, go for it. I mean, it could be contradicted by people.

37:16.960 --> 37:21.920
 Not your own instinct of based on everything you've learned. So of course you have to adapt

37:21.920 --> 37:29.440
 your beliefs when your experiments contradict those beliefs but you have to stick to your

37:29.440 --> 37:36.160
 beliefs otherwise. It's what allowed me to go through those years. It's what allowed me to

37:37.120 --> 37:44.480
 persist in directions that, you know, took time, whatever other people think, took time to mature

37:44.480 --> 37:53.680
 and bring fruits. So history of AI is marked with these, of course it's marked with technical

37:53.680 --> 37:58.880
 breakthroughs but it's also marked with these seminal events that capture the imagination

37:58.880 --> 38:06.000
 of the community. Most recent, I would say AlphaGo beating the world champion human go player

38:06.000 --> 38:14.000
 was one of those moments. What do you think the next such moment might be? Okay, sir, first of all,

38:14.000 --> 38:24.880
 I think that these so called seminal events are overrated. As I said, science really moves by

38:24.880 --> 38:33.760
 small steps. Now what happens is you make one more small step and it's like the drop that,

38:33.760 --> 38:40.560
 you know, allows to, that fills the bucket and then you have drastic consequences because now

38:40.560 --> 38:46.240
 you're able to do something you were not able to do before or now say the cost of building some

38:46.240 --> 38:51.920
 device or solving a problem becomes cheaper than what existed and you have a new market that opens

38:51.920 --> 39:00.080
 up. So especially in the world of commerce and applications, the impact of a small scientific

39:00.080 --> 39:07.520
 progress could be huge but in the science itself, I think it's very, very gradual and

39:07.520 --> 39:15.280
 where are these steps being taken now? So there's unsupervised, right? So if I look at one trend

39:15.280 --> 39:24.080
 that I like in my community, for example, and at me line, my institute, what are the two hardest

39:24.080 --> 39:32.800
 topics? GANs and reinforcement learning, even though in Montreal in particular, like reinforcement

39:32.800 --> 39:39.600
 learning was something pretty much absent just two or three years ago. So it is really a big

39:39.600 --> 39:48.400
 interest from students and there's a big interest from people like me. So I would say this is

39:48.400 --> 39:54.960
 something where we're going to see more progress even though it hasn't yet provided much in terms of

39:54.960 --> 40:01.280
 actual industrial fallout. Like even though there's Alpha Gold, there's no, like Google is not making

40:01.280 --> 40:06.320
 money on this right now. But I think over the long term, this is really, really important for many

40:06.320 --> 40:13.760
 reasons. So in other words, I would say reinforcement learning maybe more generally agent learning

40:13.760 --> 40:17.520
 because it doesn't have to be with rewards. It could be in all kinds of ways that an agent

40:17.520 --> 40:23.040
 is learning about its environment. Now, reinforcement learning, you're excited about. Do you think

40:23.040 --> 40:32.320
 GANs could provide something? Yes. Some moment in it. Well, GANs or other

40:33.760 --> 40:41.360
 generative models, I believe, will be crucial ingredients in building agents that can understand

40:41.360 --> 40:48.880
 the world. A lot of the successes in reinforcement learning in the past has been with policy

40:48.880 --> 40:53.360
 gradient where you'll just learn a policy. You don't actually learn a model of the world. But

40:53.360 --> 40:58.640
 there are lots of issues with that. And we don't know how to do model based RL right now. But I

40:58.640 --> 41:06.080
 think this is where we have to go in order to build models that can generalize faster and better,

41:06.080 --> 41:13.200
 like to new distributions that capture, to some extent, at least the underlying causal

41:13.200 --> 41:20.320
 mechanisms in the world. Last question. What made you fall in love with artificial intelligence?

41:20.960 --> 41:28.400
 If you look back, what was the first moment in your life when you were fascinated by either

41:28.400 --> 41:33.600
 the human mind or the artificial mind? You know, when I was an adolescent, I was reading a lot.

41:33.600 --> 41:41.920
 And then I started reading science fiction. There you go. That's it. That's where I got hooked.

41:41.920 --> 41:50.160
 And then, you know, I had one of the first personal computers and I got hooked in programming.

41:50.960 --> 41:55.040
 And so it just, you know, start with fiction and then make it a reality. That's right.

41:55.040 --> 42:12.080
 Yosha, thank you so much for talking to me. My pleasure.