1
00:00:07,733 --> 00:00:10,135
>> GOOD MORNING, EVERYONE.

2
00:00:10,135 --> 00:00:12,037
WELCOME TO THE THIRD TALK IN THE

3
00:00:12,037 --> 00:00:13,672
BIOWULF SERIES.

4
00:00:13,672 --> 00:00:16,008
THIS YEAR, WE ARE CELEBRATING 25

5
00:00:16,008 --> 00:00:17,409
YEARS OF PROVIDING SCIENTIFIC 

6
00:00:17,409 --> 00:00:18,811
COMPUTING TO THE NIH SCIENTIFIC 

7
00:00:18,811 --> 00:00:19,545
COMMUNITY.

8
00:00:19,545 --> 00:00:21,447
TODAY, WE ARE HONORED TO HAVE 

9
00:00:21,447 --> 00:00:23,449
DR. SAMEER ANTANI AS OUR 

10
00:00:23,449 --> 00:00:23,682
SPEAKER.

11
00:00:23,682 --> 00:00:26,118
DR. ANTANI IS A PRINCIPAL 

12
00:00:26,118 --> 00:00:26,819
INVESTIGATOR IN THE NATIONAL 

13
00:00:26,819 --> 00:00:28,721
LIBRARY OF MEDICINE.

14
00:00:28,721 --> 00:00:31,757
HE EARNED HIS PhD AND MASTER OF 

15
00:00:31,757 --> 00:00:33,859
ENGINEERING COMPUTER SCIENCE AND

16
00:00:33,859 --> 00:00:36,028
HIS B.E. IN COMPUTER 

17
00:00:36,028 --> 00:00:36,328
ENGINEERING.

18
00:00:36,328 --> 00:00:39,331
DR. ANTANI IS A FELLOW OF THE 

19
00:00:39,331 --> 00:00:41,033
AMERICAN INSTITUTE FOR MEDICAL 

20
00:00:41,033 --> 00:00:43,736
AND BIOLOGICAL ENGINEERING, 

21
00:00:43,736 --> 00:00:44,904
FELLOW OF THE INSTITUTE MUCH 

22
00:00:44,904 --> 00:00:45,471
ELECTRICAL AND ELECTRONICS 

23
00:00:45,471 --> 00:00:47,973
ENGINEER, IEEE, AND A SENIOR 

24
00:00:47,973 --> 00:00:52,278
MEMBER OF THE SPIE.

25
00:00:52,278 --> 00:00:53,879
HIS RESEARCH FOCUSES A NOVEL 

26
00:00:53,879 --> 00:00:57,850
ALGORITHMS IN MEDICAL IMAGING 

27
00:00:57,850 --> 00:00:58,484
AND MULTIMODAL MACHINE LEARNING 

28
00:00:58,484 --> 00:01:02,221
AND AI.

29
00:01:02,221 --> 00:01:05,658
AND NOW FOR GLOBAL HEALTH 

30
00:01:05,658 --> 00:01:08,160
SETTING, CHARACTERIZING DATA AND

31
00:01:08,160 --> 00:01:09,895
INNOVATIONS IN MACHINE LEARNING 

32
00:01:09,895 --> 00:01:11,630
AND ALGORITHM DESIGN FOR 

33
00:01:11,630 --> 00:01:14,433
RELIABLE AI PREDICTIONS WITH 

34
00:01:14,433 --> 00:01:17,269
APPLICATIONS IN THE SCREENING, 

35
00:01:17,269 --> 00:01:18,237
DIAGNOSTIC, RISK PREDICTION AND 

36
00:01:18,237 --> 00:01:21,240
TREATMENT.

37
00:01:21,240 --> 00:01:22,241
DR. ANTANI'S RESEARCH ON BIOWULF

38
00:01:22,241 --> 00:01:25,744
HAS IN THE PAST FOCUSED ON DEEP 

39
00:01:25,744 --> 00:01:28,914
LEARNING FOR MEDICAL IMAGE 

40
00:01:28,914 --> 00:01:30,282
IMPLEMENTATION AND RESEARCH 

41
00:01:30,282 --> 00:01:32,217
RELATED TO VARIOUS DISEASES SUCH

42
00:01:32,217 --> 00:01:34,453
AS CERVICAL CANCER AND CARDIAC 

43
00:01:34,453 --> 00:01:35,754
REJECTION FRACTION.

44
00:01:35,754 --> 00:01:40,392
ON BIOWULF, HE HAS MADE -- OTHER

45
00:01:40,392 --> 00:01:43,028
RELATED FIREWORKS.

46
00:01:43,028 --> 00:01:44,163
DR. SAMEER ANTANI'S TALK TODAY 

47
00:01:44,163 --> 00:01:48,534
IS TITLED EXPLORATIONS IN USE OF

48
00:01:48,534 --> 00:01:50,035
SYNTHETIC IMAGES FOR BIOMEDICAL 

49
00:01:50,035 --> 00:01:50,869
MONTICELLO LEARNING AND AI 

50
00:01:50,869 --> 00:01:51,136
RESEARCH.

51
00:01:51,136 --> 00:01:53,038
WELCOME.

52
00:01:53,038 --> 00:01:54,106
>> WELL, THANK YOU.

53
00:01:54,106 --> 00:01:56,775
THANK YOU VERY MUCH, AND 

54
00:01:56,775 --> 00:01:58,110
WELCOME, EVERYBODY.

55
00:01:58,110 --> 00:02:02,314
I AM GOING TO BE SPEAKING WITH 

56
00:02:02,314 --> 00:02:04,917
SYNTHESIZING IMAGES AND USING 

57
00:02:04,917 --> 00:02:06,885
THEM IN SOME APPLICATIONS 

58
00:02:06,885 --> 00:02:11,090
RELATED TO DISEASES THAT WE HAVE

59
00:02:11,090 --> 00:02:15,227
ONGOING, AND THIS IS A 

60
00:02:15,227 --> 00:02:17,463
STANDARD -- I'LL GO OVER A BRIEF

61
00:02:17,463 --> 00:02:18,630
FEW MINUTES ON WHAT KINDS OF 

62
00:02:18,630 --> 00:02:19,965
THINGS I'M WORKING ON AND THAT 

63
00:02:19,965 --> 00:02:22,368
WILL ESTABLISH THE CONTEXT 

64
00:02:22,368 --> 00:02:24,103
WITHIN WHICH SOME OF THESE 

65
00:02:24,103 --> 00:02:27,673
EFFORTS WERE EXPENDED.

66
00:02:27,673 --> 00:02:29,341
AND TRY TO LAY THE FOUNDATION OF

67
00:02:29,341 --> 00:02:33,779
HOW WE THINK ABOUT SYNTHESIZING 

68
00:02:33,779 --> 00:02:39,651
IMAGES AND TALK ABOUT A BIT OF 

69
00:02:39,651 --> 00:02:40,386
BACKGROUND IN GENERATIVE 

70
00:02:40,386 --> 00:02:43,055
ADVERSARIAL NETWORKS, LATENT 

71
00:02:43,055 --> 00:02:44,390
DIFFUSION MODELS IN PARTICULAR, 

72
00:02:44,390 --> 00:02:45,991
AND SOME OF OUR SUBPOENASES OVER

73
00:02:45,991 --> 00:02:48,560
THE PAST SIX OR SEVEN YEARS IN 

74
00:02:48,560 --> 00:02:51,730
TRYING THIS TECHNOLOGY AS SUCH.

75
00:02:51,730 --> 00:02:53,899
AND SOME CONFOUNDING RESULTS AND

76
00:02:53,899 --> 00:02:55,501
SEASONAL VERY GOOD RESULTS.

77
00:02:55,501 --> 00:02:58,404
SO THAT SHOULD HOPEFULLY LAY THE

78
00:02:58,404 --> 00:02:59,571
GROUNDWORK THAT IT IS PROMISING 

79
00:02:59,571 --> 00:03:02,775
AND VIABLE, BUT IT HAS SOME -- 

80
00:03:02,775 --> 00:03:04,043
IT LEAVES SOMETHING FOR THE USER

81
00:03:04,043 --> 00:03:06,378
OF THE TECHNOLOGY TO THINK ABOUT

82
00:03:06,378 --> 00:03:08,380
WHEN THEY ARE -- WHEN THEY'RE 

83
00:03:08,380 --> 00:03:09,248
ADOPTING THIS.

84
00:03:09,248 --> 00:03:11,850
A BIG THANKS TO MY RESEARCH TEAM

85
00:03:11,850 --> 00:03:14,286
AT THIS MOMENT.

86
00:03:14,286 --> 00:03:15,754
I HAVE SOME HERE IN THE 

87
00:03:15,754 --> 00:03:17,823
AUDIENCE, A SCIENTIST IN OUR 

88
00:03:17,823 --> 00:03:18,023
GROUP.

89
00:03:18,023 --> 00:03:23,729
A POST-DOCTORAL FELLOW, ALSO A 

90
00:03:23,729 --> 00:03:25,597
POST-DOCTOR FELLOW AND JAYLENE 

91
00:03:25,597 --> 00:03:26,832
WHO IS NOT HERE IN THE ROOM WITH

92
00:03:26,832 --> 00:03:28,600
US TODAY, BUT SHE'S MY STAFF 

93
00:03:28,600 --> 00:03:30,202
SCIENTIST AND OVER THE YEARS, 

94
00:03:30,202 --> 00:03:31,303
WE'VE HAD A NUMBER OF PEOPLE 

95
00:03:31,303 --> 00:03:32,905
CONTRIBUTE WHO ARE NOT PICTURED 

96
00:03:32,905 --> 00:03:33,338
HERE.

97
00:03:33,338 --> 00:03:35,374
BROADLY AS WITH AN INTRODUCTION,

98
00:03:35,374 --> 00:03:37,910
I CONDUCT RESEARCH IN RELIABLE 

99
00:03:37,910 --> 00:03:38,477
AND INTERPRETABLE MACHINE 

100
00:03:38,477 --> 00:03:42,281
LEARNING AND AI AND THE KEYWORD 

101
00:03:42,281 --> 00:03:44,716
HERE IS RELIABLE.

102
00:03:44,716 --> 00:03:46,285
WE'VE OBSERVED AS THE DATASETS 

103
00:03:46,285 --> 00:03:47,853
CHANGE, EVEN WITHIN THE SAME 

104
00:03:47,853 --> 00:03:49,988
TOPIC AREA, WE EXPAND TO OTHER 

105
00:03:49,988 --> 00:03:51,557
ACQUISITION MODELS OR 

106
00:03:51,557 --> 00:03:54,560
ACQUISITION METHODS, THE OUTCOME

107
00:03:54,560 --> 00:03:56,528
IS NOT NECESSARILY ANALOGOUS TO 

108
00:03:56,528 --> 00:03:57,663
WHAT WE SAW IN THE EXPERIMENT 

109
00:03:57,663 --> 00:04:00,499
TAM SECTION, BUT I BEGAN WITH.

110
00:04:00,499 --> 00:04:02,534
OR WE CHANGE THE BALANCE OF 

111
00:04:02,534 --> 00:04:04,837
CONTROLS AND WE SEE A DIFFERENT 

112
00:04:04,837 --> 00:04:06,305
BEHAVIOR.

113
00:04:06,305 --> 00:04:07,573
WE CHANGE THE CONDITIONS ON 

114
00:04:07,573 --> 00:04:09,942
IMAGES WITH THE EXPOSURE OR 

115
00:04:09,942 --> 00:04:11,009
CONTRAST ISSUE, THE BEHAVIOR 

116
00:04:11,009 --> 00:04:11,643
CHANGES AGAIN.

117
00:04:11,643 --> 00:04:15,047
SO NOW ONE COULD SAY THAT YOU 

118
00:04:15,047 --> 00:04:17,850
JUST HAVE TO DO FEWER, MAYBE 

119
00:04:17,850 --> 00:04:19,284
THAT IS TRUE, BUT THERE ARE SOME

120
00:04:19,284 --> 00:04:20,652
OBVIOUS THINGS THAT WE BELIEVE 

121
00:04:20,652 --> 00:04:22,054
THE METHOD SHOULD NOT MISS, AND 

122
00:04:22,054 --> 00:04:22,955
YET THEY DO.

123
00:04:22,955 --> 00:04:24,356
AND WITH REGARDS TO 

124
00:04:24,356 --> 00:04:25,791
INTERPRETATION, WHAT I'D LIKE TO

125
00:04:25,791 --> 00:04:27,459
ASPIRE TO DO, AND THIS IS 

126
00:04:27,459 --> 00:04:28,861
SOMETHING THAT WE ARE -- THE 

127
00:04:28,861 --> 00:04:30,796
BASIC STUFF WE DO IS 

128
00:04:30,796 --> 00:04:32,564
HIGHLIGHTING THE PARTS OF THE 

129
00:04:32,564 --> 00:04:34,600
IMAGE THAT MOST ATTENTION WAS 

130
00:04:34,600 --> 00:04:36,368
GIVEN BY THE ALGORITHMS.

131
00:04:36,368 --> 00:04:39,071
THAT TO ME IS THE BABY FIRST 

132
00:04:39,071 --> 00:04:39,271
STEP.

133
00:04:39,271 --> 00:04:40,839
I'D LIKE TO GET TO A POINT WHERE

134
00:04:40,839 --> 00:04:43,542
THEHINES ARE ABLE TO USE, 

135
00:04:43,542 --> 00:04:44,810
PARTICULARLY IN TODAY'S 

136
00:04:44,810 --> 00:04:47,980
ITERATION, OF MULTIMODAL AI, USE

137
00:04:47,980 --> 00:04:50,182
THE IDEA OF TRYING TO EXPLAIN AS

138
00:04:50,182 --> 00:04:51,850
TO WHAT THE FEATURES MEAN OR 

139
00:04:51,850 --> 00:04:54,353
WHAT THE IMAGE FEATURES MEAN, 

140
00:04:54,353 --> 00:04:56,455
WHAT DOES IT TRANSLATE INTO AND 

141
00:04:56,455 --> 00:04:57,956
USE THE CLINICAL CONTEXT TO BE 

142
00:04:57,956 --> 00:04:59,158
ABLE TO INTERPRET WHAT THE 

143
00:04:59,158 --> 00:05:00,893
OUTCOME REALLY MEANS, WHAT THE 

144
00:05:00,893 --> 00:05:01,593
PREDICTION REALLY MEANS.

145
00:05:01,593 --> 00:05:03,328
SO THAT'S WHERE I'D LIKE TO GO 

146
00:05:03,328 --> 00:05:05,864
AND THAT RELATES TO THE 

147
00:05:05,864 --> 00:05:07,132
AUGMENTATION OF THE CLINICAL AND

148
00:05:07,132 --> 00:05:07,766
RESEARCH APPLICATIONS BECAUSE 

149
00:05:07,766 --> 00:05:09,735
MUCH OF MY WORK IS IN 

150
00:05:09,735 --> 00:05:11,203
COLLABORATION WITH VARIOUS 

151
00:05:11,203 --> 00:05:14,373
INSTITUTES, HEART AND BLOOD, 

152
00:05:14,373 --> 00:05:18,544
NCI, OTHER CANCER, WITH UCSF ON 

153
00:05:18,544 --> 00:05:18,977
SARCOMA.

154
00:05:18,977 --> 00:05:21,446
IN ALL THESE -- THE ENDS POINT 

155
00:05:21,446 --> 00:05:22,614
ARE THEY'RE TRYING TO EITHER 

156
00:05:22,614 --> 00:05:24,416
DERIVE ON DECISION OF WHAT THE 

157
00:05:24,416 --> 00:05:26,952
BEST PLACE TO BIOPSY, SHOULD WE 

158
00:05:26,952 --> 00:05:28,320
BIOPSY ON ONE END AND ON THE 

159
00:05:28,320 --> 00:05:30,289
OTHER END, TRY IT IN THE FIELD 

160
00:05:30,289 --> 00:05:33,825
AND MINIMIZE THE NUMBER OF 

161
00:05:33,825 --> 00:05:37,629
ERRORS THAT POSSIBLY MIGHT BE 

162
00:05:37,629 --> 00:05:38,263
MAKING.

163
00:05:38,263 --> 00:05:40,666
SO THERE IS THIS FACTOR OF 

164
00:05:40,666 --> 00:05:42,167
TRANSLATING THE OUTCOMES OF 

165
00:05:42,167 --> 00:05:43,168
MACHINE LEARNING ALGORITHMS INTO

166
00:05:43,168 --> 00:05:44,203
THE FIELD AS WELL.

167
00:05:44,203 --> 00:05:46,371
AND THAT'S A CONTEXT FOR GLOBAL 

168
00:05:46,371 --> 00:05:49,708
HEALTH THAT I'VE MENTIONED 

169
00:05:49,708 --> 00:05:50,108
BELOW.

170
00:05:50,108 --> 00:05:51,677
SO WITH THAT, WE NEED TO START 

171
00:05:51,677 --> 00:05:53,645
THINKING ABOUT, WELL, IF YOU 

172
00:05:53,645 --> 00:05:55,814
HAVE A METHOD X THAT DOES VERY 

173
00:05:55,814 --> 00:05:57,316
WELL, THEN THE VARIABLE THAT 

174
00:05:57,316 --> 00:05:59,618
WE'RE PLAYING WITH THE MAINLY 

175
00:05:59,618 --> 00:06:02,054
DATA, AND THE DATA IS HAVING 

176
00:06:02,054 --> 00:06:05,891
SOME EFFECT, OBVIOUSLY, IN THE 

177
00:06:05,891 --> 00:06:09,027
LABORATORY IN -- IN THE 

178
00:06:09,027 --> 00:06:09,761
RELIABILITY AND GENERALIZATION 

179
00:06:09,761 --> 00:06:11,096
OF ALGORITHMS AND WE'VE BROKEN 

180
00:06:11,096 --> 00:06:12,564
THEM DOWN TO VARIOUS FACTORS 

181
00:06:12,564 --> 00:06:13,966
DEPENDING ON WHAT WE GET FROM 

182
00:06:13,966 --> 00:06:15,534
THESE INSTITUTES AND 

183
00:06:15,534 --> 00:06:16,535
COLLABORATORS AND SITES 

184
00:06:16,535 --> 00:06:21,273
WORLDWIDE, IS THE CONCEPT OF 

185
00:06:21,273 --> 00:06:21,573
QUALITY.

186
00:06:21,573 --> 00:06:25,377
THERE'S EXPOSURE, THERE'S BLUR, 

187
00:06:25,377 --> 00:06:28,847
THERE IS SHADE, THERE'S CONTACT 

188
00:06:28,847 --> 00:06:31,683
ISSUES, BUT THE FRAMING OF THE 

189
00:06:31,683 --> 00:06:33,151
ORGAN OF INTEREST, REGION OF 

190
00:06:33,151 --> 00:06:34,886
INTEREST WAS NOT ACCURATE, AND 

191
00:06:34,886 --> 00:06:35,520
SO ON.

192
00:06:35,520 --> 00:06:37,823
SO THERE'S A SPECTRUM OF THINGS.

193
00:06:37,823 --> 00:06:40,459
THERE ARE ALSO OTHER MUNDANE 

194
00:06:40,459 --> 00:06:42,594
ISSUES LIKE YOU HAVE THE 

195
00:06:42,594 --> 00:06:44,529
PHOTO -- IF YOUR PATIENT'S BY 

196
00:06:44,529 --> 00:06:45,530
THE FOLDER AND YOU HAVE THE 

197
00:06:45,530 --> 00:06:47,032
WRONG IMAGE IN THE WRONG FOLDER.

198
00:06:47,032 --> 00:06:49,935
IT SOUNDS SILLY, BUT THERE'S -- 

199
00:06:49,935 --> 00:06:51,670
SO THE QUALITY SPECTRUM IS VERY 

200
00:06:51,670 --> 00:06:51,970
BROAD.

201
00:06:51,970 --> 00:06:54,640
IT GOES INTO ACTUAL PROCESSING 

202
00:06:54,640 --> 00:06:56,875
QUALITY TO DATA QUALITY IN 

203
00:06:56,875 --> 00:06:58,844
GENERAL.

204
00:06:58,844 --> 00:07:01,780
THE VOLUME IS WELL UNDERSTOOD 

205
00:07:01,780 --> 00:07:01,980
ISSUE.

206
00:07:01,980 --> 00:07:03,348
THAT'S NOT ALWAYS TRUE AS SOME 

207
00:07:03,348 --> 00:07:04,783
OF THE OUTCOMES SHOW BECAUSE WE 

208
00:07:04,783 --> 00:07:07,986
TRIED TO COLLECT A NUMBER OF 

209
00:07:07,986 --> 00:07:10,722
IMAGES -- THE NUMBER OF UNITS OF

210
00:07:10,722 --> 00:07:12,024
DATA THAT THE ALGORITHM 

211
00:07:12,024 --> 00:07:12,424
PROCESSING WITH.

212
00:07:12,424 --> 00:07:14,259
AND THAT'S NOT ALWAYS TRUE.

213
00:07:14,259 --> 00:07:15,627
MORE IS NOT BETTER.

214
00:07:15,627 --> 00:07:17,462
MAYBE DIVERSE IS BETTER AND THAT

215
00:07:17,462 --> 00:07:20,032
CAN LEAD TO VARIETY, BUT WHAT DO

216
00:07:20,032 --> 00:07:20,966
YOU MEAN BY DIVERSITY?

217
00:07:20,966 --> 00:07:23,669
IS IT JUST A CASE OF THE NUMBER 

218
00:07:23,669 --> 00:07:25,704
OF CASES YOU'VE SEEN, WHAT KIND 

219
00:07:25,704 --> 00:07:27,673
OF REGIONS ARE YOU SEEING IN THE

220
00:07:27,673 --> 00:07:30,509
IMAGE, ARE WE USING ANATOMICAL 

221
00:07:30,509 --> 00:07:31,943
KNOWLEDGE OR PATHOLOGICAL 

222
00:07:31,943 --> 00:07:34,112
KNOWLEDGE TO INFORM THE MACHINE 

223
00:07:34,112 --> 00:07:34,680
LEARNING ALGORITHMS?

224
00:07:34,680 --> 00:07:38,050
DO WE HAVE ENOUGH VARIETY IN THE

225
00:07:38,050 --> 00:07:42,220
PATHOLOGY TYPES OR IF THERE ARE 

226
00:07:42,220 --> 00:07:43,655
THINGS THAT ARE PRESENT, THE 

227
00:07:43,655 --> 00:07:45,057
SEVERITY OF DISEASE AND WHAT 

228
00:07:45,057 --> 00:07:46,291
POINT YOU CATCH IT.

229
00:07:46,291 --> 00:07:47,893
PARTICULARLY YOU GET INTO 

230
00:07:47,893 --> 00:07:49,594
LONGITUDINAL DATASETS AND THIS 

231
00:07:49,594 --> 00:07:52,064
PREDICTION AND THIS MATTERS A 

232
00:07:52,064 --> 00:07:52,931
LOT.

233
00:07:52,931 --> 00:07:54,866
HOW DO YOU HANDLE IMBALANCE AND 

234
00:07:54,866 --> 00:07:56,301
WHAT FACTOR IT DOES PLAY?

235
00:07:56,301 --> 00:07:58,937
THEN YOU COME TO THE HUMAN 

236
00:07:58,937 --> 00:08:01,039
ELEMENT OF THIS.

237
00:08:01,039 --> 00:08:04,242
IN SUPERVISED LEARNING, WE ASK 

238
00:08:04,242 --> 00:08:07,279
EXPERTS TO ANOTE TATE THE IMAGES

239
00:08:07,279 --> 00:08:09,448
FOR US AS WHERE THEY BELIEVE THE

240
00:08:09,448 --> 00:08:10,749
DISEASE IS WHAN THE MACHINE 

241
00:08:10,749 --> 00:08:11,216
SHOULD LEARN.

242
00:08:11,216 --> 00:08:12,484
WE'VE FOUND GREAT VARIETY IN THE

243
00:08:12,484 --> 00:08:13,218
EXPERTS THEMSELVES.

244
00:08:13,218 --> 00:08:15,020
HAT THEY'RE WRONG, IT'S 

245
00:08:15,020 --> 00:08:16,488
THAT THEIR PERSPECTIVE ON THE 

246
00:08:16,488 --> 00:08:18,757
IMMEDIACY OF CARE, AND 

247
00:08:18,757 --> 00:08:20,125
THEREFORE, RATING THE PATIENT AS

248
00:08:20,125 --> 00:08:23,829
HIGH GRADE OR MEDIUM, MODEST OR 

249
00:08:23,829 --> 00:08:26,331
LOW GRADE, VARIES BASED ON WHAT 

250
00:08:26,331 --> 00:08:27,432
THEY PERCEIVE OR THEIR 

251
00:08:27,432 --> 00:08:28,300
EXPERIENCE TELLS THEM SHOULD BE 

252
00:08:28,300 --> 00:08:30,001
SOMETHING THAT THE MACHINE 

253
00:08:30,001 --> 00:08:32,571
SHOULD PAY ATTENTION TO.

254
00:08:32,571 --> 00:08:35,340
AND A SIMPLE EXAMPLE WE HAVE 

255
00:08:35,340 --> 00:08:37,142
FROM COVID WAS WE HAD, 

256
00:08:37,142 --> 00:08:38,076
IMMEDIATELY AFTER AND IN THE 

257
00:08:38,076 --> 00:08:39,644
MIDST OF COVID, A BUNCH OF 

258
00:08:39,644 --> 00:08:42,314
IMPRASE THAT CAME AROUND AND WE 

259
00:08:42,314 --> 00:08:44,483
ASKED TWO RADIOLOGISTS OF RENOWN

260
00:08:44,483 --> 00:08:47,185
TO TRY AND ANNOTATE THEM FOR US.

261
00:08:47,185 --> 00:08:48,220
ONE WORKED IN A FIELD HOSPITAL 

262
00:08:48,220 --> 00:08:50,422
AND THE OTHER WAS A RESEARCHER.

263
00:08:50,422 --> 00:08:52,591
THE RESEARCHER WAS FAR MORE 

264
00:08:52,591 --> 00:08:54,826
METICULOUS, FAR MORE SELECTIVE, 

265
00:08:54,826 --> 00:08:56,228
VERY CONSERVATIVE BECAUSE HE -- 

266
00:08:56,228 --> 00:08:59,398
I GUESS THE MINDSET WAS, I WANT 

267
00:08:59,398 --> 00:09:02,200
TO BE MORE ACCURATE, WHILE THE 

268
00:09:02,200 --> 00:09:03,735
FIELD GUY WAS -- THIS REGION IS 

269
00:09:03,735 --> 00:09:04,002
FINE.

270
00:09:04,002 --> 00:09:05,804
PICK THE ENTIRE REGION AND GO 

271
00:09:05,804 --> 00:09:07,839
WITH IT.

272
00:09:07,839 --> 00:09:12,944
SO YOU HAVE THESE -- THESE 

273
00:09:12,944 --> 00:09:14,713
FIGURES ARE COMPLETELY OFF, 

274
00:09:14,713 --> 00:09:16,681
TRYING TO ARRIVE AT SOME KIND OF

275
00:09:16,681 --> 00:09:18,850
CONSENSUS AND THAT IS A FACTOR.

276
00:09:18,850 --> 00:09:20,485
HOW DOES LONGITUDINAL DATA PLAY 

277
00:09:20,485 --> 00:09:21,853
A ROLE AND IF YOU HAVE SAME 

278
00:09:21,853 --> 00:09:23,488
PATIENT OVER TIME, ACCOUNTING 

279
00:09:23,488 --> 00:09:25,190
AND ACCOMMODATING FOR THE AMOUNT

280
00:09:25,190 --> 00:09:27,025
OF MEMORY THAT THE MODEL HAS, 

281
00:09:27,025 --> 00:09:30,662
HOW YOU COMPENSATE FOR THAT.

282
00:09:30,662 --> 00:09:33,799
AND USE OF TABULAR DATA OR ANY 

283
00:09:33,799 --> 00:09:36,668
OTHER TEXT THAT SURROUNDS THE 

284
00:09:36,668 --> 00:09:37,769
IMAGES THAT COULD BE VALUABLE 

285
00:09:37,769 --> 00:09:39,037
FOR PROVIDING LABELS.

286
00:09:39,037 --> 00:09:41,139
AND WHAT OF UNCERTAINTY?

287
00:09:41,139 --> 00:09:44,309
WE CLASSIFY CONFIDENCE AS A 

288
00:09:44,309 --> 00:09:45,877
PROBABILITY OF THE ELEMENT 

289
00:09:45,877 --> 00:09:47,279
BELONGING TO A PARTICULAR CLASS,

290
00:09:47,279 --> 00:09:49,047
BUT IS IT TRULY PROBABILITY?

291
00:09:49,047 --> 00:09:51,817
IS IT PROBABILITY THAT WITHIN 

292
00:09:51,817 --> 00:09:53,418
THAT YOU ARE EXPOSED TO, WHICH 

293
00:09:53,418 --> 00:09:54,519
COMES BACK TO THE VOLUME OF DATA

294
00:09:54,519 --> 00:09:56,621
YOU'RE TALKING GOOD, OR IS IT IN

295
00:09:56,621 --> 00:09:58,156
GENERAL THAT NO MATTER WHAT, 

296
00:09:58,156 --> 00:09:59,324
THIS IS THE INCIDENCE RATE AND 

297
00:09:59,324 --> 00:10:02,360
HOW DOES IT RELATE TO THE EP 

298
00:10:02,360 --> 00:10:03,862
PEOPLOLOGY PART OF IT.

299
00:10:03,862 --> 00:10:06,364
-- EPIDEMIOLOGY PART OF IT.

300
00:10:06,364 --> 00:10:07,933
BIAS, LOOKING AT DARK SKINNED 

301
00:10:07,933 --> 00:10:09,501
VERSUS LIGHT SKINNED IMAGES.

302
00:10:09,501 --> 00:10:11,436
MOST OF THE IMAGES WE HAVE ARE 

303
00:10:11,436 --> 00:10:12,671
OF LIGHTER SKIN TONES AND IF 

304
00:10:12,671 --> 00:10:14,773
YOU'RE PARTICULARLY LOOKING FOR 

305
00:10:14,773 --> 00:10:17,008
DARK SKIN TONES DISEASE 

306
00:10:17,008 --> 00:10:19,678
PATTERNS, SARCOMA BEING A BROWN 

307
00:10:19,678 --> 00:10:21,446
PATCH OF PRESENTATION TYPES ON 

308
00:10:21,446 --> 00:10:23,782
VERY DARK, UGANDAN SKIN OR SOUTH

309
00:10:23,782 --> 00:10:27,686
AFRICAN SKIN, THEN YOU DON'T 

310
00:10:27,686 --> 00:10:29,087
HAVE THE DATA THAT WOULD 

311
00:10:29,087 --> 00:10:31,990
ACTUALLY BUILD A FOUNDATION  FL.

312
00:10:31,990 --> 00:10:33,358
THOSE ARE FACTORS THAT PLAY INTO

313
00:10:33,358 --> 00:10:34,893
THE ROLE OF MY WORK AND THIS IS 

314
00:10:34,893 --> 00:10:36,962
JUST A QUICK SCREEN SHOT OF SOME

315
00:10:36,962 --> 00:10:38,797
OF THE THINGS WE TALKED ABOUT, 

316
00:10:38,797 --> 00:10:41,266
SO I WON'T BELABOR THAT TOO 

317
00:10:41,266 --> 00:10:41,466
MUCH.

318
00:10:41,466 --> 00:10:43,435
SO BASICALLY, THAT SORT OF 

319
00:10:43,435 --> 00:10:45,203
BUILDS THE CASE FOR WHY 

320
00:10:45,203 --> 00:10:47,239
SYNTHESIZING DATA.

321
00:10:47,239 --> 00:10:51,109
WE ARE COLLECTING FOR KAPOSI AND

322
00:10:51,109 --> 00:10:53,211
REDUCING VISCOSITY BYSI REDUCIN 

323
00:10:53,211 --> 00:10:54,579
THE VOLUME OF NUMBER OF SAMPLES.

324
00:10:54,579 --> 00:10:56,314
HOPING TO INCREASE THE VARIETY, 

325
00:10:56,314 --> 00:10:58,216
AT LEAST IN TERMS OF CASES.

326
00:10:58,216 --> 00:11:00,218
IMAGE REGION IS IN THERE BECAUSE

327
00:11:00,218 --> 00:11:04,155
WE NOTE WHEN WE COLLECTED DATA 

328
00:11:04,155 --> 00:11:06,057
TO CERVICAL CANCER FROM VARIOUS 

329
00:11:06,057 --> 00:11:07,692
SITES, TO THE HUMAN EYE, IT HADE

330
00:11:07,692 --> 00:11:08,593
NO DIFFERENCE.

331
00:11:08,593 --> 00:11:10,428
TO THE MACHINE, WE COULD CLUSTER

332
00:11:10,428 --> 00:11:13,164
THEM IN SEPARATE CLUSTERS, WHICH

333
00:11:13,164 --> 00:11:13,865
WAS BIZARRE.

334
00:11:13,865 --> 00:11:15,934
SO THE QUESTION, IF WE ARE DID 

335
00:11:15,934 --> 00:11:17,302
GO TO TRY AND ROLL THIS OUT IN 

336
00:11:17,302 --> 00:11:18,803
THE FIELD, THEN DOES IT MAKE 

337
00:11:18,803 --> 00:11:22,674
SENSE TO TRY AND ARRIVE AT 

338
00:11:22,674 --> 00:11:24,042
MOVING IMAGES FROM ONE CLUSTER 

339
00:11:24,042 --> 00:11:26,478
TO ANOTHER SO THAT YOUR MODEL 

340
00:11:26,478 --> 00:11:29,447
ACTUALLY STARTS BEHAVING BETTER?

341
00:11:29,447 --> 00:11:31,550
BIAS WE TALKED ABOUT AND 

342
00:11:31,550 --> 00:11:34,986
IMBALANCE AS WELL IN THE PAST.

343
00:11:34,986 --> 00:11:35,587
SO THERE ARE DIFFERENT KINDS 

344
00:11:35,587 --> 00:11:36,988
OF -- AS MOST PEOPLE WHO ARE IN 

345
00:11:36,988 --> 00:11:39,224
THE FIELD MAY KNOW AND THIS IS 

346
00:11:39,224 --> 00:11:40,358
NOT SUPPOSED TO BE EXHAUSTIVE, 

347
00:11:40,358 --> 00:11:41,927
NOR IS IT SUPPOSED TO BE 

348
00:11:41,927 --> 00:11:43,161
COMPLETELY A TALK THAT TELLS YOU

349
00:11:43,161 --> 00:11:45,964
HOW TO DO THESE THINGS, BUT 

350
00:11:45,964 --> 00:11:50,035
ABOUT THE KINDS OF -- THEY ALL 

351
00:11:50,035 --> 00:11:53,872
BEGAN VARIATIONAL ENCODERS THAT 

352
00:11:53,872 --> 00:11:56,441
CAN BE USED TO MAKE NEW DATA AND

353
00:11:56,441 --> 00:11:57,642
TRY TO DIVERSE THE FEATURESSEOU 

354
00:11:57,642 --> 00:12:00,111
HAVE COMPRESSED INTO A SMALLER 

355
00:12:00,111 --> 00:12:00,412
SPACE.

356
00:12:00,412 --> 00:12:03,181
AND AN EXAMPLE IS SHOWN IN THE 

357
00:12:03,181 --> 00:12:04,549
FIGURE WHERE YOU HAVE, DID YOU 

358
00:12:04,549 --> 00:12:06,585
HAVE AN ENCODER THAT COMPRESSES 

359
00:12:06,585 --> 00:12:11,456
THE FEATURES AND DECODER 

360
00:12:11,456 --> 00:12:13,725
RECREATES X, VERY CLOSE TO X 

361
00:12:13,725 --> 00:12:14,059
ITSELF.

362
00:12:14,059 --> 00:12:15,260
AND THEREFORE, YOU HAVE THOSE 

363
00:12:15,260 --> 00:12:16,928
FEATURES D, YOU CAN HOPEFULLY 

364
00:12:16,928 --> 00:12:18,229
CREATE NEW VARIATIONS.

365
00:12:18,229 --> 00:12:21,733
THAT WAS THE ORIGINAL IDEA.

366
00:12:21,733 --> 00:12:24,102
YOU GET INTO AUTO REGRESSIVE 

367
00:12:24,102 --> 00:12:25,437
MODELS, THE FOUNDATIONAL ELEMENT

368
00:12:25,437 --> 00:12:27,272
OF LARGE LANGUAGE MODELS TODAY 

369
00:12:27,272 --> 00:12:29,107
IN GPT-3 AND THEY ARE 

370
00:12:29,107 --> 00:12:31,109
ESSENTIALLY TRANSFORM MERS.

371
00:12:31,109 --> 00:12:32,644
THEY ARE MORE SEQUENTIAL 

372
00:12:32,644 --> 00:12:34,279
PREDICTORS OF TOKENS APPEARING 

373
00:12:34,279 --> 00:12:37,582
IN THE CONTEXT OF A LARGER 

374
00:12:37,582 --> 00:12:39,117
FRAGMENT OF DATA, AND IN THE 

375
00:12:39,117 --> 00:12:40,585
IMAGING YOU CAN THINK OF THEM AS

376
00:12:40,585 --> 00:12:41,152
SEQUENTIAL IMAGES.

377
00:12:41,152 --> 00:12:42,854
YOU CAN ALSO THINK OF THEM AS 

378
00:12:42,854 --> 00:12:44,155
PART OF THE IMAGES THAT ARE 

379
00:12:44,155 --> 00:12:46,124
RELATED TO EACH OTHER THAT 

380
00:12:46,124 --> 00:12:47,192
CO-OCCUR AND AS YOU'RE WALKING 

381
00:12:47,192 --> 00:12:49,427
THROUGH AN IMAGE, YOU WILL 

382
00:12:49,427 --> 00:12:51,896
ESSENTIALLY CREATE A MODEL THAT 

383
00:12:51,896 --> 00:12:55,233
WOULD USE THAT TO YOUR 

384
00:12:55,233 --> 00:12:56,601
ADVANTAGE.

385
00:12:56,601 --> 00:13:00,972
I MOVED THEM DOWN HERE, THEY'RE 

386
00:13:00,972 --> 00:13:03,475
ACTUALLY IN TIME POINT RIGHT 

387
00:13:03,475 --> 00:13:05,977
AFTER ENCODERS, WHERE QUOUF TWO 

388
00:13:05,977 --> 00:13:07,646
COMPETING NETWORKS, ONE THAT IS 

389
00:13:07,646 --> 00:13:08,980
SYNTHESIZING POSHLY FOR 

390
00:13:08,980 --> 00:13:10,882
RANDOM -- POTENTIALLY FOR RANDOM

391
00:13:10,882 --> 00:13:12,751
NOISE, BUT YOU CAN REPLACE THE 

392
00:13:12,751 --> 00:13:14,252
RANDOM NOISE BLOCK -- OOPS, 

393
00:13:14,252 --> 00:13:16,187
SORRY, BACK AGAIN ONE.

394
00:13:16,187 --> 00:13:17,889
YOU HAVE THAT RANDOMIZED BLOCK 

395
00:13:17,889 --> 00:13:19,090
RIGHT THERE, YOU CAN CHANGE IT 

396
00:13:19,090 --> 00:13:20,892
WITH ANOTHER MODEL THAT IS 

397
00:13:20,892 --> 00:13:22,060
TRAINED OFF OF AN IMAGE THAT YOU

398
00:13:22,060 --> 00:13:23,461
HAVE SITTING WITH YOU, AND YOU 

399
00:13:23,461 --> 00:13:26,631
CAN START FROM A POINT THAT'S 

400
00:13:26,631 --> 00:13:28,667
CLOSER TO WHAT YOU WANT THE 

401
00:13:28,667 --> 00:13:30,201
DISCRIMINATOR TO DISCRIMINATE 

402
00:13:30,201 --> 00:13:30,435
AGAINST.

403
00:13:30,435 --> 00:13:31,770
AND THIS IS ANOTHER MODEL THAT 

404
00:13:31,770 --> 00:13:33,138
IS JUST LEARNING OVER TIME TO 

405
00:13:33,138 --> 00:13:35,473
DECIDE IF THE SYNTHESIZED MODEL 

406
00:13:35,473 --> 00:13:38,009
IS CLOSE ENOUGH TO WHAT IT WOULD

407
00:13:38,009 --> 00:13:39,778
CONSIDER AS BEING IN CLASS 

408
00:13:39,778 --> 00:13:42,480
VERSUS OUT OF CLASS.

409
00:13:42,480 --> 00:13:44,182
AND SO THE PROPERTY OF GANNS, 

410
00:13:44,182 --> 00:13:47,485
THEREFORE, UNLIKE THE DIFFUSION 

411
00:13:47,485 --> 00:13:49,621
MODELS THAT COMES LATER IS 

412
00:13:49,621 --> 00:13:51,790
CREATING LOOK-ALIKES THAT BELONG

413
00:13:51,790 --> 00:13:53,091
WITH SMALL VARIATIONS THAT ARE 

414
00:13:53,091 --> 00:13:54,192
PER CHANCE USUALLY.

415
00:13:54,192 --> 00:13:57,262
I'M GOING TO SKIP THE FLOW-BASED

416
00:13:57,262 --> 00:13:59,531
MODEL FOR COMPLETENESS, BUT THE 

417
00:13:59,531 --> 00:14:02,333
LATENT DIFFUSION MODELS ARE 

418
00:14:02,333 --> 00:14:05,503
STILL -- IT'S A BRANDED PRODUCT.

419
00:14:05,503 --> 00:14:08,506
BASICALLY TRIES TO GO ONE STEP 

420
00:14:08,506 --> 00:14:11,276
BEYOND THAT WHERE THE DIFFUSION 

421
00:14:11,276 --> 00:14:12,677
PROCESS ADDS MORE NOISE AT 

422
00:14:12,677 --> 00:14:14,179
VARYING SCALES AND VARYING 

423
00:14:14,179 --> 00:14:18,049
DEGREES OF NOISINESS TO AN 

424
00:14:18,049 --> 00:14:21,252
IMAGE, THAT THE MODEL TRIES TO 

425
00:14:21,252 --> 00:14:23,455
LEARN AND RECOVER THOSE IMAGES 

426
00:14:23,455 --> 00:14:23,722
FROM.

427
00:14:23,722 --> 00:14:26,391
AND THE IDEA THEREFORE IS THAT 

428
00:14:26,391 --> 00:14:27,826
IF YOU HAVE AN IMAGE GOING IN 

429
00:14:27,826 --> 00:14:30,128
WITH ALL THE NOISE IN THERE AND 

430
00:14:30,128 --> 00:14:32,230
IT'S TRYING TO UNDO THE NOISE, 

431
00:14:32,230 --> 00:14:34,566
IT ALSO IS IMPLICITLY LEARNING 

432
00:14:34,566 --> 00:14:35,333
THE IMAGE FEATURES THAT ARE 

433
00:14:35,333 --> 00:14:37,001
TRULY BELONGING TO THE IMAGE 

434
00:14:37,001 --> 00:14:38,470
BECAUSE EACH TIME THERE'S A 

435
00:14:38,470 --> 00:14:42,307
LOSS, IT HAS LEARNED IT MADE A 

436
00:14:42,307 --> 00:14:44,142
MISTAKE AND WHAT WAS ACHIEVED 

437
00:14:44,142 --> 00:14:45,376
THROUGH NOISE AND WHAT WAS 

438
00:14:45,376 --> 00:14:47,545
ACHIEVED THROUGH CONTENT.

439
00:14:47,545 --> 00:14:48,346
IT BECOMES A LITTLE MORE 

440
00:14:48,346 --> 00:14:50,648
INTERESTING WHEN YOU START 

441
00:14:50,648 --> 00:14:51,316
ADDING TEXT.

442
00:14:51,316 --> 00:14:52,016
OKAY.

443
00:14:52,016 --> 00:14:53,852
I'LL GET TO THAT.

444
00:14:53,852 --> 00:14:55,386
ADDING TEXT TO THAT WHERE YOU 

445
00:14:55,386 --> 00:14:58,256
NOW HAVE TEXT LABELS GOING ALONG

446
00:14:58,256 --> 00:15:00,725
WITH THE IMAGE AND THE CONTEXT 

447
00:15:00,725 --> 00:15:02,694
OF THOSE PARTS OF THE IMAGE THAT

448
00:15:02,694 --> 00:15:05,897
IT'S LEARNING AS TRUE IMAGE AND 

449
00:15:05,897 --> 00:15:10,135
NOT BEING NOISE ARE NOW BETTER 

450
00:15:10,135 --> 00:15:11,903
ASSOCIATED AND YOU HAVE A 

451
00:15:11,903 --> 00:15:13,104
PAIR-WISE CONTRAST OF LEARNING 

452
00:15:13,104 --> 00:15:14,773
GOING ON WHERE YOU HAVE A TEXT 

453
00:15:14,773 --> 00:15:17,242
LABEL OR TEXT CONCEPT COMBINED 

454
00:15:17,242 --> 00:15:18,943
WITH IMAGE FEATURES ALONG WITH 

455
00:15:18,943 --> 00:15:20,645
NOISE AT VARIOUS SCALES, AND 

456
00:15:20,645 --> 00:15:23,348
ENOUGH DATA, IT TENDS TO NOW 

457
00:15:23,348 --> 00:15:25,150
SYNTHESIZE WHAT YOU WANT BASED 

458
00:15:25,150 --> 00:15:26,951
ON THE CONCEPT THAT YOU 

459
00:15:26,951 --> 00:15:28,553
REQUESTED, ASSUMING THAT THE 

460
00:15:28,553 --> 00:15:29,754
CONCEPT WAS THERE IN THE FIRST 

461
00:15:29,754 --> 00:15:31,456
PLACE IN THE TRAINING MODEL, SO 

462
00:15:31,456 --> 00:15:34,793
THAT CONDITION STILL REMAINS.

463
00:15:34,793 --> 00:15:37,028
SO I'VE ALREADY DESCRIBED THIS 

464
00:15:37,028 --> 00:15:40,465
GOING THROUGH THE GAN ELEMENT, 

465
00:15:40,465 --> 00:15:42,767
AND THE KIND OF GAN, WHAT YOU 

466
00:15:42,767 --> 00:15:45,036
SAW IN THE PREVIOUS SLIDE THERE,

467
00:15:45,036 --> 00:15:46,704
BASICALLY AGAIN, YOU HAVE NOISE 

468
00:15:46,704 --> 00:15:48,273
AND YOUR DISCRIMINATOR AND IT'S 

469
00:15:48,273 --> 00:15:49,607
BEEN TRAINED TO SOMETHING AND 

470
00:15:49,607 --> 00:15:53,144
KEEP ON LEARNING UNTIL IT FITS 

471
00:15:53,144 --> 00:15:55,380
THE LAST FUNCTION.

472
00:15:55,380 --> 00:15:57,949
BUT THERE ARE VARIATIONS IN 

473
00:15:57,949 --> 00:15:58,149
THIS.

474
00:15:58,149 --> 00:15:59,284
YOU CAN DECIDE TO SCALE IMAGES 

475
00:15:59,284 --> 00:16:01,352
UP AND DOWN AND IT WILL LEARN 

476
00:16:01,352 --> 00:16:03,655
MORE AND CAN USE THE INDIVIDUAL 

477
00:16:03,655 --> 00:16:05,957
LAYERS NOW TO LEARN MORE ABOUT 

478
00:16:05,957 --> 00:16:06,958
THE FEATURES THEMSELVES.

479
00:16:06,958 --> 00:16:11,296
AND SO YOUR CONTEXT IS NOW 

480
00:16:11,296 --> 00:16:12,764
ESTABLISHED BETTER BY 

481
00:16:12,764 --> 00:16:14,632
ACCOMMODATING FOR IMAGES 

482
00:16:14,632 --> 00:16:16,034
APPEARINGAPG OR LOWER 

483
00:16:16,034 --> 00:16:17,368
RESOLUTION THAT MAY BE -- SO 

484
00:16:17,368 --> 00:16:19,838
AGAIN, SELECTING WHICH KIND OF 

485
00:16:19,838 --> 00:16:20,972
GAN YOU WANT FOR YOUR 

486
00:16:20,972 --> 00:16:23,208
APPLICATION BECOMES AN IMP A

487
00:16:23,208 --> 00:16:23,474
FACTOR.

488
00:16:23,474 --> 00:16:26,644
CONDITION GAN NOW APPLIES -- 

489
00:16:26,644 --> 00:16:29,180
USES LABELS TO SAY CERTAIN 

490
00:16:29,180 --> 00:16:29,914
IMAGES, CERTAIN IMAGE FEATURES 

491
00:16:29,914 --> 00:16:31,916
CAN ONLY CO-OCCUR IF SOMETHING 

492
00:16:31,916 --> 00:16:34,152
ELSE HAS HAPPENED AND YOU BUILT 

493
00:16:34,152 --> 00:16:38,857
THAT INTO THE GAN ITSELF.

494
00:16:38,857 --> 00:16:42,293
GAN IS LOT LIKE THE VANILLA GAN,

495
00:16:42,293 --> 00:16:44,362
EXCEPT THERE ARE DIFFERENT LOSS 

496
00:16:44,362 --> 00:16:46,564
FUNCTIONS FOR IMAGE APPEARANCE 

497
00:16:46,564 --> 00:16:46,798
REGIONS.

498
00:16:46,798 --> 00:16:47,599
YOU WANT SOMETHING TO REALLY 

499
00:16:47,599 --> 00:16:48,600
LOOK LIKE SOMETHING AND WE HAVE 

500
00:16:48,600 --> 00:16:49,934
AN EXAMPLE IN THE SLIDES COMING 

501
00:16:49,934 --> 00:16:52,337
AHEAD WHERE WE USE THAT TO 

502
00:16:52,337 --> 00:16:57,075
REMOVE FOCAL BLUR IN IMAGES.

503
00:16:57,075 --> 00:17:00,945
FOR IMAGE -- YOU HAVE CYCLE GAN 

504
00:17:00,945 --> 00:17:03,214
THAT IS A PAIR-WISE MATCHING 

505
00:17:03,214 --> 00:17:04,949
BETWEEN TEXT CONCEPTS WHERE TWO 

506
00:17:04,949 --> 00:17:06,351
IDEAS AND THEN IT WILL TRANSLATE

507
00:17:06,351 --> 00:17:10,021
FROM ONE TO THE OTHER, AND STYLE

508
00:17:10,021 --> 00:17:13,057
IS CONTROLLING, AS SUGGESTS, THE

509
00:17:13,057 --> 00:17:15,193
APPEARANCE OF AN IMAGE FOR A 

510
00:17:15,193 --> 00:17:16,227
PARTICULAR SCALE -- IF YOU WERE 

511
00:17:16,227 --> 00:17:17,862
TO SYNTHESIZE A PAINTING IN THE 

512
00:17:17,862 --> 00:17:19,731
STYLE OF PICASSO, PERHAPS, AND 

513
00:17:19,731 --> 00:17:22,100
YOU'RE NOT LOOKING A LATENT 

514
00:17:22,100 --> 00:17:26,037
DIFFUSION MODEL, YOU COULD 

515
00:17:26,037 --> 00:17:28,139
EXTRACT T WAY CERTAIN PATTERNS

516
00:17:28,139 --> 00:17:29,173
CO-OCCUR ON THE IMAGES AND YOU 

517
00:17:29,173 --> 00:17:30,942
HAVE A LABEL THAT'S ASSIGNED TO 

518
00:17:30,942 --> 00:17:32,076
THAT.

519
00:17:32,076 --> 00:17:34,512
SO THESE ARE DIFFERENT KINDS OF 

520
00:17:34,512 --> 00:17:35,914
GANs THAT ONE SHOULD THINK ABOUT

521
00:17:35,914 --> 00:17:37,148
AND THERE ARE A FEW OTHERS THAT 

522
00:17:37,148 --> 00:17:38,850
YOU CAN THINK ABOUT IN TERMS OF 

523
00:17:38,850 --> 00:17:40,184
WHAT YOUR APPLICATION REALLY 

524
00:17:40,184 --> 00:17:40,485
DEMANDS.

525
00:17:40,485 --> 00:17:42,687
THAT'S WHERE THE NUANCE STARTS 

526
00:17:42,687 --> 00:17:45,857
CREEPING IN TO THIS DISCUSSION.

527
00:17:45,857 --> 00:17:52,797
I WENT OVER THIS IN THE PAST, 

528
00:17:52,797 --> 00:17:56,668
AND GOING TO REMOVE NOISE IN A 

529
00:17:56,668 --> 00:17:57,502
STEP-BY-STEP FASHION IN EACH 

530
00:17:57,502 --> 00:17:59,003
ITERATION AS WE REMOVE THE NOISE

531
00:17:59,003 --> 00:18:00,204
ARE IT LEARNS SOMETHING MORE 

532
00:18:00,204 --> 00:18:01,739
ABOUT THE IMAGE.

533
00:18:01,739 --> 00:18:03,708
THE MORE FEATURES IN EAC S

534
00:18:03,708 --> 00:18:05,710
OF NOISE REMOVAL IS ALSO 

535
00:18:05,710 --> 00:18:06,878
EXTRACTING FEATURES AND IF YOU 

536
00:18:06,878 --> 00:18:07,345
COMBINE THEM WITH

537
00:18:07,345 --> 00:18:10,548
TEXT, YOU GET 

538
00:18:10,548 --> 00:18:13,584
A BETTER MATCHING.

539
00:18:13,584 --> 00:18:17,388
AND, OF COURSE, GAN, WHILE THESE

540
00:18:17,388 --> 00:18:20,158
ITERATIONS AT DIFFERENT SCALES, 

541
00:18:20,158 --> 00:18:24,128
IT'S GOT MORE POWER AND MORE 

542
00:18:24,128 --> 00:18:24,729
FLEXIBILITY TH FOR THE 

543
00:18:24,729 --> 00:18:26,597
KINDS OF DATA THAT TR NOT 

544
00:18:26,597 --> 00:18:27,665
NECESSARILY A SINGLE FRAME 

545
00:18:27,665 --> 00:18:29,734
IMAGE, BUT TEMPORAL DATA OR 

546
00:18:29,734 --> 00:18:32,103
ASSOCIATED FRAMES OR SLICES IN A

547
00:18:32,103 --> 00:18:33,571
CT THAT ARE CORRELATED WILL 

548
00:18:33,571 --> 00:18:35,573
START MAKING A LOT MORE SENSE TO

549
00:18:35,573 --> 00:18:37,542
THE MODEL ITSELF AS OPPOSED TO A

550
00:18:37,542 --> 00:18:39,911
SINGLE X-RAY THAT'S GOING 

551
00:18:39,911 --> 00:18:40,244
THROUGH.

552
00:18:40,244 --> 00:18:42,547
SO BOTH -- SO IT OFFERS MUCH 

553
00:18:42,547 --> 00:18:44,015
MORE POWER AND CONSEQUENTIAL 

554
00:18:44,015 --> 00:18:45,717
DEMANDS MUCH MORE COMPUTE, SO 

555
00:18:45,717 --> 00:18:47,819
THAT'S THE OTHER COST C THAT C  

556
00:18:47,819 --> 00:18:49,287
HAVE TO PAY.

557
00:18:49,287 --> 00:18:50,521
AND THIS IS IN THE CLIP PART, 

558
00:18:50,521 --> 00:18:53,825
WHICH YOU'RE ADDING TEXT AND 

559
00:18:53,825 --> 00:18:58,696
IMAGE TOGETHER, AND FOR IMAGE 

560
00:18:58,696 --> 00:19:00,431
CAPTURING REASONS OR REVERSE, 

561
00:19:00,431 --> 00:19:02,300
REQUESTING IMAGES THAT BELONG TO

562
00:19:02,300 --> 00:19:05,370
A PARTICULAR CLASS BY HAVING A 

563
00:19:05,370 --> 00:19:10,208
TEXT CAPTION THAT GOES WITH IT, 

564
00:19:10,208 --> 00:19:11,376
IDENTIFYING OBJECTS AND IMAGES, 

565
00:19:11,376 --> 00:19:13,177
ASSUME YOU HAVE A LARGE ENOUGH 

566
00:19:13,177 --> 00:19:14,178
SET AND YOU HAVE THE WORD 

567
00:19:14,178 --> 00:19:15,480
ASSOCIATED WITH THE CONTEXT OF 

568
00:19:15,480 --> 00:19:16,948
IMAGES IN THE ASSOCIATED TEXT 

569
00:19:16,948 --> 00:19:18,783
THAT GOES WITH THAT IMAGE, THEN 

570
00:19:18,783 --> 00:19:19,517
FINDING THOSE OBJECTS IS NOW 

571
00:19:19,517 --> 00:19:22,120
THEREFORE DERIVABLE AND IT IS 

572
00:19:22,120 --> 00:19:24,389
ESSENTIALLY REPETITION OF A 

573
00:19:24,389 --> 00:19:25,523
PATTERN OVER TIME THAT THE MAGIC

574
00:19:25,523 --> 00:19:28,126
REALLY HAPPENS.

575
00:19:28,126 --> 00:19:30,194
AND TEXT TO IMAGE, WE'RE DOING 

576
00:19:30,194 --> 00:19:33,131
SOME TEXT-TO-IMAGE WORK AS WELL 

577
00:19:33,131 --> 00:19:34,098
AND IMAGE CLASSIFICATION WORK AS

578
00:19:34,098 --> 00:19:38,603
WELL IN SOME OF OUR MORE RECENT 

579
00:19:38,603 --> 00:19:39,103
APPROACHES.

580
00:19:39,103 --> 00:19:44,842
SO A VARIANT OF MODELS, WHICH IS

581
00:19:44,842 --> 00:19:46,044
ABOUT CLASSIFYING PRESERVATION, 

582
00:19:46,044 --> 00:19:46,944
WHERE BECAUSE YOU'RE BEING IN 

583
00:19:46,944 --> 00:19:49,313
THE MEDICAL DOMAIN, WE ARE 

584
00:19:49,313 --> 00:19:50,948
STARTING WITH PRETRAINED MODELS 

585
00:19:50,948 --> 00:19:53,618
THAT HAVE BEEN BUILT OFF OF 

586
00:19:53,618 --> 00:19:55,887
STOCK PHOTOGRAPHY.

587
00:19:55,887 --> 00:19:59,557
NOW, WHILE THEY WORK, IT'S NOT 

588
00:19:59,557 --> 00:20:01,025
OPTIMAL AND ESPECIALLY IN -- 

589
00:20:01,025 --> 00:20:02,860
YOU'LL SEE SOME EXAMPLES AHEAD 

590
00:20:02,860 --> 00:20:05,997
WHERE YOU SEE SOME -- I MEAN 

591
00:20:05,997 --> 00:20:09,634
TODAY, IF YOU WENT TO DAL I OR 

592
00:20:09,634 --> 00:20:13,237
GEMINI OR OPENING EYES VERSION 

593
00:20:13,237 --> 00:20:15,973
OF THAT, YOU HAVE AN X-RAY AND 

594
00:20:15,973 --> 00:20:18,876
YOU SEE GLOSSY NON-LIFE LIKE 

595
00:20:18,876 --> 00:20:19,544
SILHOUETTES WHICH IS OBVIOUSLY 

596
00:20:19,544 --> 00:20:21,646
WRONG, AND IT'S BECAUSE IT 

597
00:20:21,646 --> 00:20:23,815
DOESN'T HAVE THE DATA, ALL THE 

598
00:20:23,815 --> 00:20:26,918
IMAGE PIXEL EXPRESSION TO 

599
00:20:26,918 --> 00:20:29,687
SYNTHESIZE THOSE KIND OF IMAGES.

600
00:20:29,687 --> 00:20:32,223
SO A CLASS SPECIFIC FEATURE 

601
00:20:32,223 --> 00:20:33,524
PRESERVATION ALLOWS YOU TO 

602
00:20:33,524 --> 00:20:34,792
PRODUCE YOUR OWN IMAGES ON TOP 

603
00:20:34,792 --> 00:20:35,960
AND ADD YOUR OWN LABELS THAT 

604
00:20:35,960 --> 00:20:38,062
MAKE SORE SENSE, BUT STILL BUILD

605
00:20:38,062 --> 00:20:39,230
OFF OF THE MODEL.

606
00:20:39,230 --> 00:20:40,465
YOU'RE IN THE BUILDING A 

607
00:20:40,465 --> 00:20:42,300
FOUNDATIONAL MODEL, SO WE DON'T 

608
00:20:42,300 --> 00:20:43,768
NEED TO HAVE TWO MILLION OR 

609
00:20:43,768 --> 00:20:45,103
THREE MILLION X-RAYS TO START 

610
00:20:45,103 --> 00:20:46,104
DOING AN X-RAY MODEL.

611
00:20:46,104 --> 00:20:48,773
YOU CAN START WITH THOUSANDS, 

612
00:20:48,773 --> 00:20:50,975
BUT YOU ARE SORT OF FINE TUNING 

613
00:20:50,975 --> 00:20:52,977
OR SHAPING IT WITH YOUR CLASSES 

614
00:20:52,977 --> 00:20:54,145
THAT ARE IMPORTANT TO YOUR NEEDS

615
00:20:54,145 --> 00:20:56,647
AND YOUR APPLICATIONS.

616
00:20:56,647 --> 00:21:02,854
TO RECONSTRUCT THE IMAGE AS HIGH

617
00:21:02,854 --> 00:21:03,087
QUALITY.

618
00:21:03,087 --> 00:21:09,527
NOW, WHILE THIS ALL SOUNDS VERY 

619
00:21:09,527 --> 00:21:11,496
LIKE THAT IT WORKS LIKE MAGIC, 

620
00:21:11,496 --> 00:21:12,730
THERE ARE LOTS OF CAUTIONS HERE 

621
00:21:12,730 --> 00:21:14,198
BECAUSE IT DOESN'T UNDERSTAND 

622
00:21:14,198 --> 00:21:16,334
DISEASE, IT DOESN'T UNDERSTAND 

623
00:21:16,334 --> 00:21:17,668
ANATOMY, IT DOESN'T -- OBSERVING

624
00:21:17,668 --> 00:21:19,570
WHAT WE'VE BEEN DOING, I GO BACK

625
00:21:19,570 --> 00:21:24,909
TO THE END OF MY TALK WITH -- 

626
00:21:24,909 --> 00:21:26,711
LITTLE KIDS ARE BEING X-RAYED, 

627
00:21:26,711 --> 00:21:28,346
THEIR MOMS ARE THERE, THEY SE,

628
00:21:28,346 --> 00:21:30,414
AROUND AND SO THERE'S NO 

629
00:21:30,414 --> 00:21:30,948
CONSISTENCY.

630
00:21:30,948 --> 00:21:33,284
SO IT BECOMES A -- TO US, WE NOW

631
00:21:33,284 --> 00:21:35,653
START LOCATING THE REGIONS OF 

632
00:21:35,653 --> 00:21:38,756
INTEREST THAT WE WANTED TO FOCUS

633
00:21:38,756 --> 00:21:41,359
ON AND CORRECT FOR THE 

634
00:21:41,359 --> 00:21:43,227
ORIENTATION, AND THEN IT WORKS 

635
00:21:43,227 --> 00:21:45,463
IN THAT SUBCONTEXT.

636
00:21:45,463 --> 00:21:47,665
SO THERE'S -- HOW MUCH OF THIS 

637
00:21:47,665 --> 00:21:49,800
IS ART, HOW MUCH OF THIS IS 

638
00:21:49,800 --> 00:21:50,668
ENGINEERING, AND HOURCH OF THIS 

639
00:21:50,668 --> 00:21:52,904
IS RECOGNIZING THE CLINICAL 

640
00:21:52,904 --> 00:21:54,772
CONTEXT ARE VERY IMPORTANT 

641
00:21:54,772 --> 00:21:59,277
FACTORS IN THE WAY YOU DESIGN 

642
00:21:59,277 --> 00:22:01,445
AND IT'S USUALLY NOT ONE 

643
00:22:01,445 --> 00:22:03,948
MONOLITH, IT IS A PROCESS THAT 

644
00:22:03,948 --> 00:22:04,782
YOU HAVE TO SEQUENCE THROUGH.

645
00:22:04,782 --> 00:22:06,717
THAT'S ANOTHER FACTOR THAT'S 

646
00:22:06,717 --> 00:22:07,351
IMPORTANTS ABOUT THIS IS SOME 

647
00:22:07,351 --> 00:22:08,452
WORK WE'VE DONE RECENTLY WHERE 

648
00:22:08,452 --> 00:22:10,488
WE'RE GOING TO SYNTHESIZE X-RAY 

649
00:22:10,488 --> 00:22:10,721
IMAGES.

650
00:22:10,721 --> 00:22:13,224
I THINK THIS IS SANDY FROM 

651
00:22:13,224 --> 00:22:19,697
OUR -- WE ARE PARTNERING IN AN 

652
00:22:19,697 --> 00:22:22,300
ACQUISITION CHALLENGE ORGANIZED 

653
00:22:22,300 --> 00:22:24,535
WHERE THEY WANT TO GO BEYOND 

654
00:22:24,535 --> 00:22:26,771
STATE OF THE ART, NOT TO SAY YOU

655
00:22:26,771 --> 00:22:28,940
HAVE PNEUMONIA, YOU DON'T HAVE 

656
00:22:28,940 --> 00:22:30,074
KNEW MOAN YEAH, BUT TO ACTUALLY 

657
00:22:30,074 --> 00:22:31,175
SAY WHICH PART OF THE LUNG 

658
00:22:31,175 --> 00:22:34,212
ACTUALLY HAS IT AND HOW MUCH IS 

659
00:22:34,212 --> 00:22:34,412
THERE.

660
00:22:34,412 --> 00:22:35,846
SO THEY'RE TALKING ABOUT MORE OF

661
00:22:35,846 --> 00:22:37,048
SPECIFIC ABOUT THE EXPRESSION OF

662
00:22:37,048 --> 00:22:37,281
DISEASE.

663
00:22:37,281 --> 00:22:38,749
AND AS USUAL, WE DON'T HAVE 

664
00:22:38,749 --> 00:22:40,384
ENOUGH SAMPLES THAT ARE OF THAT 

665
00:22:40,384 --> 00:22:44,889
KIND, SO WE WANTED TO START WITH

666
00:22:44,889 --> 00:22:45,990
EXISTING IMAGES OUT IN THE OPEN 

667
00:22:45,990 --> 00:22:48,259
AND USE THEM TO BUILD MORE 

668
00:22:48,259 --> 00:22:50,127
SPECIFIC MEASUREMENTS THAT 

669
00:22:50,127 --> 00:22:55,333
ARE -- NOW, FID AND KID IN THE 

670
00:22:55,333 --> 00:22:58,636
COLUMNS THERE REFER TO THE 

671
00:22:58,636 --> 00:22:59,337
PERCEPTION DISTANCE AND I'LL 

672
00:22:59,337 --> 00:23:01,105
LEAVE THAT TO YOU TO LOOK UP, 

673
00:23:01,105 --> 00:23:02,073
BUT BASICALLY LOWER IS BETTER.

674
00:23:02,073 --> 00:23:04,442
HOW FAR IS THE SYNTHESIZED IMAGE

675
00:23:04,442 --> 00:23:06,711
FROM THE ORIGINAL IMAGE IN TERMS

676
00:23:06,711 --> 00:23:08,079
OF FUTURE SPACE AND THERE ARE 

677
00:23:08,079 --> 00:23:10,648
TWO DIFFERENT WAYS OF LOOKING AT

678
00:23:10,648 --> 00:23:11,449
IT.

679
00:23:11,449 --> 00:23:15,753
AND AS YOU CAN SEE WITH PRIOR 

680
00:23:15,753 --> 00:23:16,587
PRESERVATION, IT'S LOWER 

681
00:23:16,587 --> 00:23:17,922
COMPARED TO A IMPAN AND OTHER 

682
00:23:17,922 --> 00:23:21,292
DIFFUSION MODELS THAT TRIES 

683
00:23:21,292 --> 00:23:23,427
TO -- AND FORGIVE ME FOR THE 

684
00:23:23,427 --> 00:23:25,363
FLIP OF THE BEFORE AND AFTER 

685
00:23:25,363 --> 00:23:28,899
TUNING, BUT THIS IS WHAT CAME 

686
00:23:28,899 --> 00:23:32,203
OFF JUST PLAIN VANILLA OUT OF 

687
00:23:32,203 --> 00:23:33,437
THE BASIC MODEL AND YOU CAN SEE 

688
00:23:33,437 --> 00:23:34,839
IT'S NOTHING LIKE AN X-RAY THAT 

689
00:23:34,839 --> 00:23:36,007
YOU WOULD BELIEVE IS AN X-RAY, 

690
00:23:36,007 --> 00:23:38,376
AND THIS IS WHAT WE ENDED UP 

691
00:23:38,376 --> 00:23:38,576
WITH.

692
00:23:38,576 --> 00:23:39,944
WHILE ON THIS SIDE, THIS IS WHAT

693
00:23:39,944 --> 00:23:41,746
WE END UP WITH, WHICH IS MUCH 

694
00:23:41,746 --> 00:23:44,181
LIKE WHAT WE WANT AND THIS IS 

695
00:23:44,181 --> 00:23:45,816
NOT EVEN CT.

696
00:23:45,816 --> 00:23:46,517
I DON'T KNOW WHAT THIS IS.

697
00:23:46,517 --> 00:23:50,221
SO CLEARLY, WHAT THIS TELLS YOU 

698
00:23:50,221 --> 00:23:53,524
IS THAT IF YOU DOWNLOAD 

699
00:23:53,524 --> 00:23:55,993
SOMETHING, BEWARE AND THEN BE 

700
00:23:55,993 --> 00:23:57,928
AWARE ABOUT WHAT TO -- OR HOW TO

701
00:23:57,928 --> 00:24:03,634
SHAPE IT TO YOUR SPECIFIC NEEDS.

702
00:24:03,634 --> 00:24:05,436
THESE ARE MORE EXAMPLES WITH 

703
00:24:05,436 --> 00:24:07,138
DISEASES OF DIFFERENT KINDS, 

704
00:24:07,138 --> 00:24:14,645
WITH COVID, WITHOUT COVID, AND A

705
00:24:14,645 --> 00:24:16,414
DIFFUSION MODEL ON THE FIRST ROW

706
00:24:16,414 --> 00:24:17,248
WHICH GIVES MORE CONTRAST 

707
00:24:17,248 --> 00:24:17,848
IMAGES.

708
00:24:17,848 --> 00:24:20,251
SO IN THIS, THE ARGUMENT CAN BE 

709
00:24:20,251 --> 00:24:22,320
MADE THAT IT'S UP TO THE EXPERT 

710
00:24:22,320 --> 00:24:26,190
EYE TO SAY THAT THIS IS NOTHING 

711
00:24:26,190 --> 00:24:30,328
LIKE HERE, OR THERE ARE MISTAKES

712
00:24:30,328 --> 00:24:30,528
HERE.

713
00:24:30,528 --> 00:24:32,263
TO THE NAIVE EYE, THEY ALL LOOK 

714
00:24:32,263 --> 00:24:33,864
PRETTY GOOD, BUT THEN WHAT IS 

715
00:24:33,864 --> 00:24:36,367
GOOD AND MAYBE COME BACK TO THE 

716
00:24:36,367 --> 00:24:38,669
QUALITY QUESTION WE COVERED, AND

717
00:24:38,669 --> 00:24:41,906
IT IIMES BEYOND THE SCOPE

718
00:24:41,906 --> 00:24:44,175
OF THE FIELD IN THE LAB TO BE 

719
00:24:44,175 --> 00:24:45,810
ABLE TO DETERMINE THAT, BUT THEN

720
00:24:45,810 --> 00:24:48,612
RELY ON A CLASSIFIER THAT HAS 

721
00:24:48,612 --> 00:24:50,414
BEEN TRAINED TO REAL DATA TO SAY

722
00:24:50,414 --> 00:24:51,682
HOW WRONG OR RIGHT WE WERE WITH 

723
00:24:51,682 --> 00:24:53,484
THE SYNTHESIZED IMAGES BEING A 

724
00:24:53,484 --> 00:24:58,556
SEPARATE TEST OBJECT.

725
00:24:58,556 --> 00:25:01,726
SOME EXAMPLES WITH X-RAYS, AN 

726
00:25:01,726 --> 00:25:04,862
INTERESTING CASE HERE, WHICH IS 

727
00:25:04,862 --> 00:25:11,702
A DISEASE OF SOME KIND AND -- SO

728
00:25:11,702 --> 00:25:13,371
THERE WAS AN INTERESTING TALK 

729
00:25:13,371 --> 00:25:14,638
YESTERDAY OR TWO DAYS AGO WHERE 

730
00:25:14,638 --> 00:25:18,142
THE SPEAKER WAS TRYING TO USE 

731
00:25:18,142 --> 00:25:20,077
SPECIES A SIGNALS ON SYNTHESIZED

732
00:25:20,077 --> 00:25:21,512
IMAGES TO SEE HOW DIFFERENT THEY

733
00:25:21,512 --> 00:25:24,248
ARE AND THIS IS AN EXAMPLE 

734
00:25:24,248 --> 00:25:26,317
WHERE, BECAUSE YOU'RE USING LDM 

735
00:25:26,317 --> 00:25:28,652
DIFFUSION MODELS, YOU'RE NOT 

736
00:25:28,652 --> 00:25:29,720
GOING TO CREATE CLONES.

737
00:25:29,720 --> 00:25:30,654
YOU'RE TRYING TO CREATE IMAGES 

738
00:25:30,654 --> 00:25:33,591
THAT BELONG TO THE SAME CLASS AS

739
00:25:33,591 --> 00:25:36,794
UNDERSTOOD BY AN EXPERT, A HUMAN

740
00:25:36,794 --> 00:25:37,094
EXPERT.

741
00:25:37,094 --> 00:25:39,597
AND THEREIN LIES THE RUB IS 

742
00:25:39,597 --> 00:25:41,399
WHERE -- WHAT IS THE LOSS 

743
00:25:41,399 --> 00:25:45,336
FUNCTION AS RELATES TO A 

744
00:25:45,336 --> 00:25:46,303
MATHEMATICAL MODEL AND WHAT IS 

745
00:25:46,303 --> 00:25:47,805
THE LOSS OF THE ERROR AS RELATES

746
00:25:47,805 --> 00:25:51,208
TO THE CONTEXT AS ESTABLISHED BY

747
00:25:51,208 --> 00:25:53,411
THE HUMAN JUDGE BECAUSE THE 

748
00:25:53,411 --> 00:25:55,079
IMAGE THAT YOU'RE EXPECTING THE 

749
00:25:55,079 --> 00:25:56,614
AI MODEL TO GIVE A 

750
00:25:56,614 --> 00:25:59,083
MEANINGFULNESS AT THE END, AND 

751
00:25:59,083 --> 00:26:02,386
THAT'S NOT ALWAYS THE CASE AND 

752
00:26:02,386 --> 00:26:05,589
SOME ARE JUST TERRIBLE IN 

753
00:26:05,589 --> 00:26:05,956
CONTRAST.

754
00:26:05,956 --> 00:26:07,625
SOME ARE MORE BELIEVABLE.

755
00:26:07,625 --> 00:26:10,628
THESE ARE NOT -- THIS IS NOT AN 

756
00:26:10,628 --> 00:26:12,363
EXHAUSTIVE LIST, BUT I PUT DOWN 

757
00:26:12,363 --> 00:26:16,233
SOME KEY ONES.

758
00:26:16,233 --> 00:26:20,871
THERE'S A FEW MISSING BETWEEN 21

759
00:26:20,871 --> 00:26:23,674
AND 22 WORK WK THE GROUP AT PENN

760
00:26:23,674 --> 00:26:25,342
STATE ON CREATING HISSO 

761
00:26:25,342 --> 00:26:26,644
PATHOLOGY IMAGES AND THOSE ARE 

762
00:26:26,644 --> 00:26:27,912
MUCH MORE TRACTABLE BECAUSE 

763
00:26:27,912 --> 00:26:30,247
THERE'S A VERY PRONOUNCED 

764
00:26:30,247 --> 00:26:30,915
STRUCTURE.

765
00:26:30,915 --> 00:26:33,451
THE CELLES FROM THE BIOPSY 

766
00:26:33,451 --> 00:26:36,487
AND THE LAYERS THERE MATTER.

767
00:26:36,487 --> 00:26:38,255
THE CELL POSITION MATTERS AND 

768
00:26:38,255 --> 00:26:39,857
THOSE STRUCTURES ARE REPEATABLE,

769
00:26:39,857 --> 00:26:42,526
SOT DID A MUCH BETTER JOB AND 

770
00:26:42,526 --> 00:26:43,727
THEY WEREERISHED, SO I DIDN'T

771
00:26:43,727 --> 00:26:44,528
INCLUDE THEM HERE.

772
00:26:44,528 --> 00:26:46,931
I PICK THE ONES THAT ARE 

773
00:26:46,931 --> 00:26:49,533
HIGHLIGHTED IN BOLD HERE TO MAKE

774
00:26:49,533 --> 00:26:51,969
A POINT ABOUT THE SUCCESSES AND 

775
00:26:51,969 --> 00:26:53,437
SOMETHING THAT MADE OUR HEADS 

776
00:26:53,437 --> 00:26:54,905
CLINK AND GO, REALLY?

777
00:26:54,905 --> 00:26:55,940
WHAT'S GOING ON HERE?

778
00:26:55,940 --> 00:26:58,442
AND WE DON'T HAVE GOOD ANSWERS.

779
00:26:58,442 --> 00:27:00,678
WE HAVE HYPOTHESES, BUT NOT GOOD

780
00:27:00,678 --> 00:27:03,113
ANSWERS FOR THAT YET.

781
00:27:03,113 --> 00:27:03,581
SO T

782
00:27:03,581 --> 00:27:07,051
 FIRST EXAMPLE WITH DEEP 

783
00:27:07,051 --> 00:27:08,752
LEARNING IMAGES, THE BACKGROUND 

784
00:27:08,752 --> 00:27:09,920
IS THE CERVICAL CANCER WORK.

785
00:27:09,920 --> 00:27:11,689
THE IDEA IS THAT A WOMAN GOES TO

786
00:27:11,689 --> 00:27:13,858
A HAVE IT, SHE'S TREATED WITH 

787
00:27:13,858 --> 00:27:16,527
THE 5% ACETIC ACID AND IF SHE 

788
00:27:16,527 --> 00:27:27,071
HAS AN HPV INFECTION, THE CERVIX

789
00:27:28,138 --> 00:27:29,640
WOULD LIGHTEN.

790
00:27:29,640 --> 00:27:33,210
AND WHAT ARE THE LESIONS?

791
00:27:33,210 --> 00:27:36,847
AND WE WANT THE GUIDANCE FOR THE

792
00:27:36,847 --> 00:27:39,583
CLINICIAN, IS THIS WOMAN -- ARE 

793
00:27:39,583 --> 00:27:41,085
THESE LESIONS CONSISTENT WITH 

794
00:27:41,085 --> 00:27:43,587
THOSE THAT ARE SEEN ON 

795
00:27:43,587 --> 00:27:45,256
PRECANCEROUS WOMEN, AND TWO, 

796
00:27:45,256 --> 00:27:47,057
WHICH ONE SHOULD WE BIOPSY AND 

797
00:27:47,057 --> 00:27:51,128
FOR A WOMAN THAT IS POSITIVE, IS

798
00:27:51,128 --> 00:27:52,563
SHE ELIGIBLE FOR ABLATION.

799
00:27:52,563 --> 00:27:53,898
THAT'S ALL THE END GOAL.

800
00:27:53,898 --> 00:27:57,635
NONE OF THIS IS BEING DESCRIBED 

801
00:27:57,635 --> 00:27:59,203
HERE.

802
00:27:59,203 --> 00:28:00,804
IN U.S. AND MOST OF THE FIRST 

803
00:28:00,804 --> 00:28:04,275
WORLD, IF I MAY USE THAT TERM, 

804
00:28:04,275 --> 00:28:07,378
IT'S A CAMERA THAT'S MOUNTED ON 

805
00:28:07,378 --> 00:28:09,680
A TRIPOD OR HEAVY WHEELS, HEAVY 

806
00:28:09,680 --> 00:28:12,049
BASE, NICE THREE ALREADY JOINTED

807
00:28:12,049 --> 00:28:17,855
ARM AND YOU GET ROCK SOLID 

808
00:28:17,855 --> 00:28:20,291
IMAGES SO THAT PROBLEM DOESN'T 

809
00:28:20,291 --> 00:28:20,824
EXIST.

810
00:28:20,824 --> 00:28:22,192
MUCH OF THE PART OF THE WORLD 

811
00:28:22,192 --> 00:28:24,194
WHERE THIS DISEASE IS HIGHLY 

812
00:28:24,194 --> 00:28:26,196
PREFER VENT AND MORE PROBLEM -- 

813
00:28:26,196 --> 00:28:27,264
PREVALENT AND MORE PROBLEMATIC 

814
00:28:27,264 --> 00:28:30,234
AND NURSES ARE GOING INTO THE 

815
00:28:30,234 --> 00:28:32,369
VILLAGES USING HANDHELD MOBILE 

816
00:28:32,369 --> 00:28:35,205
PHONES, AND THIS DEVICE, THEY 

817
00:28:35,205 --> 00:28:36,941
HAVE A PROBLEM OF SHAKE, 

818
00:28:36,941 --> 00:28:39,143
ILLUMINATION, ETCETERA.

819
00:28:39,143 --> 00:28:43,113
SO WE BEGAN TACKLING BOTH 

820
00:28:43,113 --> 00:28:45,849
ILLUMINATION AND FOCUS AND 

821
00:28:45,849 --> 00:28:47,017
MOTION BLUR AT THE SAME TIME 

822
00:28:47,017 --> 00:28:49,119
BECAUSE MOBILE PHONES WITH LOW 

823
00:28:49,119 --> 00:28:50,421
ILLUMINATION WILL TEND TO OPEN 

824
00:28:50,421 --> 00:28:53,424
THE SHUTTER FOR LONGER AND YOU 

825
00:28:53,424 --> 00:28:56,360
GET A LONGER EXPOSED IMAGE AND 

826
00:28:56,360 --> 00:28:57,828
YOU'RE NOT LOOKING AT FRAMING AT

827
00:28:57,828 --> 00:28:59,196
THIS POINT, OR WE WERE NOT AT 

828
00:28:59,196 --> 00:29:00,998
THAT POINT IN THIS WORK.

829
00:29:00,998 --> 00:29:03,634
THE IDEA WAS SIMPLE IN SOME 

830
00:29:03,634 --> 00:29:04,068
SENSE.

831
00:29:04,068 --> 00:29:05,135
YOU HAVE PAIR-WISE IMAGES, YOU 

832
00:29:05,135 --> 00:29:06,570
TRAIN IT TO KNOW WHAT IS GOOD 

833
00:29:06,570 --> 00:29:08,439
AND WHAT IS BAD AND HAVE IT 

834
00:29:08,439 --> 00:29:10,741
CORRECT FOR WHAT IS -- SO THE 

835
00:29:10,741 --> 00:29:12,576
OUTPUT OF THIS, THERE WERE, IS 

836
00:29:12,576 --> 00:29:14,144
GOING TO BE VERY -- A GAN IS THE

837
00:29:14,144 --> 00:29:15,346
BEST IDEA HERE BECAUSE YOU'RE 

838
00:29:15,346 --> 00:29:16,046
CREATING SOMETHING THAT'S 

839
00:29:16,046 --> 00:29:20,651
EXACTLY THE SAME, JUST SHORTSTOP

840
00:29:20,651 --> 00:29:24,622
R SHARPENED, SO YOU'RE IN 

841
00:29:24,622 --> 00:29:25,255
ESSENCE DEBLURRING.

842
00:29:25,255 --> 00:29:28,058
SO FOR THAT WE USE TWO KINDS OF 

843
00:29:28,058 --> 00:29:29,026
GANs, CONDITIONED ON THE 

844
00:29:29,026 --> 00:29:30,861
APPEARANCE THE IMAGE.

845
00:29:30,861 --> 00:29:32,796
WE WANTED THAT SHARPNESS OF THE 

846
00:29:32,796 --> 00:29:34,365
PAIR TO GET INTO THE FEATURE 

847
00:29:34,365 --> 00:29:37,334
EXTRACTION PROCESS AND A 

848
00:29:37,334 --> 00:29:38,168
DISTANCE MEASURE WAS USED FOR 

849
00:29:38,168 --> 00:29:40,604
THE DISCRIMINATION PROCESS, SO 

850
00:29:40,604 --> 00:29:42,106
WE WANTED IT TO HAVE THE BEST 

851
00:29:42,106 --> 00:29:46,410
FIT TO THE CLASS THAT YOU WERE 

852
00:29:46,410 --> 00:29:49,446
SYNTHESIZING THE IMAGES FOR FROM

853
00:29:49,446 --> 00:29:51,849
THE GENERATOR.

854
00:29:51,849 --> 00:29:54,118
AND THE LOSS IS NOT ANYTHING TO 

855
00:29:54,118 --> 00:29:55,686
SPEAK ABOUT.

856
00:29:55,686 --> 00:30:00,257
AND WE GOT A SET OF 2400 SHARP 

857
00:30:00,257 --> 00:30:02,126
AND 1200 BLURRY IMAGES FOR THIS 

858
00:30:02,126 --> 00:30:02,559
WORK.

859
00:30:02,559 --> 00:30:04,294
NOW, THE QUESTION BECOMES, IF 

860
00:30:04,294 --> 00:30:05,996
THE BLURRY AND THE SHARP AREN'T 

861
00:30:05,996 --> 00:30:07,765
A PAIR, YOU REALLY CAN'T DO THIS

862
00:30:07,765 --> 00:30:08,165
WORK.

863
00:30:08,165 --> 00:30:11,535
SO WE HAVE A BLUR THE SHOT ONCE 

864
00:30:11,535 --> 00:30:12,670
AND THE KINDS OF THINGS WE WERE 

865
00:30:12,670 --> 00:30:15,105
SEEING AND THEN LEARN TO UNBLUR,

866
00:30:15,105 --> 00:30:17,574
THE EXACT SAME THING THAT LDM 

867
00:30:17,574 --> 00:30:20,878
WOULD DO TODAY IS JUST NOW -- 

868
00:30:20,878 --> 00:30:23,013
THIS IS PRE-LDM DAYS, SO NOW 

869
00:30:23,013 --> 00:30:25,749
THAT WE LOOK AT 2019, OH YEAH, 

870
00:30:25,749 --> 00:30:29,386
THEY WERE USING THE SAME STUFF.

871
00:30:29,386 --> 00:30:39,930
AND WE HAVE AN 85/15 SPLIT, AND 

872
00:30:40,464 --> 00:30:43,701
THE 85 INCLUDES THE SUBSET THAT 

873
00:30:43,701 --> 00:30:45,636
ARE NOT MENTIONED HERE.

874
00:30:45,636 --> 00:30:47,237
YOU END UP WITH -- YOU WANT TO 

875
00:30:47,237 --> 00:30:49,206
MEASURE SOMETHING THAT IS OF 

876
00:30:49,206 --> 00:30:51,241
GOOD QUALITY, FIX THE NOISE 

877
00:30:51,241 --> 00:30:53,510
RATIO, WHICH IS LOWER IN THE 

878
00:30:53,510 --> 00:30:57,181
CASE OF THESE IMAGES THAN THE 

879
00:30:57,181 --> 00:30:59,450
OTHER, AND SUCH A SIMILARITY 

880
00:30:59,450 --> 00:31:04,088
MEASURE WHICH, AGAIN, LOWER ISS.

881
00:31:04,088 --> 00:31:05,989
DOESN'T MEAN MUCH, BUT IT'S A 

882
00:31:05,989 --> 00:31:07,424
VE FIT TO THE ORIGINAL 

883
00:31:07,424 --> 00:31:11,295
IMAGE.

884
00:31:11,295 --> 00:31:13,030
THE POINT OF THE GOVERNMENT ROLE

885
00:31:13,030 --> 00:31:15,733
WAS USE THE DEBLURRED IMAGES IN 

886
00:31:15,733 --> 00:31:17,868
THE CLASS FIRE TRAINED ON THE 

887
00:31:17,868 --> 00:31:18,569
NONBLURRY IMAGES.

888
00:31:18,569 --> 00:31:21,071
AND WE SAW A BOOST.

889
00:31:21,071 --> 00:31:24,408
WE SAID THERE WAS A MIX SENT 

890
00:31:24,408 --> 00:31:25,509
THROUGH WHICH HADRO PERFORM

891
00:31:25,509 --> 00:31:27,945
OF 20% LESS THAN WHAT WE WERE --

892
00:31:27,945 --> 00:31:29,246
I DON'T HAVE THE DETAILS IN MY 

893
00:31:29,246 --> 00:31:30,514
SLIDE HERE, BUT THE PAPER HAS 

894
00:31:30,514 --> 00:31:31,315
THE DETAILS.

895
00:31:31,315 --> 00:31:32,516
AND YOU SAW A BOOST.

896
00:31:32,516 --> 00:31:35,919
THE POINT WAS THAT THE 

897
00:31:35,919 --> 00:31:38,122
DEBLURRING THROUGH A GAN 

898
00:31:38,122 --> 00:31:39,323
MECHANISM DOES WORK AND DOES 

899
00:31:39,323 --> 00:31:41,959
MAKE A VIABLE OPTION FOR 

900
00:31:41,959 --> 00:31:43,260
IMPROVING FOR THE APPLICATION 

901
00:31:43,260 --> 00:31:45,763
THAT YOU'RE INTERESTED IN.

902
00:31:45,763 --> 00:31:48,832
SO THERE ARE LOTS OF FOLLOW-ONS 

903
00:31:48,832 --> 00:31:49,600
HERE.

904
00:31:49,600 --> 00:31:52,269
THEN WE WERE THINKING ABOUT HOW 

905
00:31:52,269 --> 00:31:56,273
TO ASSESS -- BOTH SPECIFIC 

906
00:31:56,273 --> 00:31:57,608
COLLECTIONS DONE FOR SPECIFIC 

907
00:31:57,608 --> 00:31:59,076
STUDIES ARE -- YOU GET IN 

908
00:31:59,076 --> 00:32:01,512
HUNDREDS LOW HUNDREDS IF YOU'RE

909
00:32:01,512 --> 00:32:01,712
LUCKY.

910
00:32:01,712 --> 00:32:02,846
ALSO, THE INCIDENCE RATE OF 

911
00:32:02,846 --> 00:32:03,680
DISEASE MATTERS.

912
00:32:03,680 --> 00:32:05,916
SO IF YOUR INCIDENCE RATE IS 

913
00:32:05,916 --> 00:32:08,418
1.5% OR 1.3% AND YOU COLLECT A 

914
00:32:08,418 --> 00:32:11,555
THOUSAND PEOPLE, YOU'RE GOING TO

915
00:32:11,555 --> 00:32:13,624
GET 150 SAMPLES.

916
00:32:13,624 --> 00:32:14,958
SO HIGH END BALANCE, YOU HAVE 

917
00:32:14,958 --> 00:32:18,328
VERY FEW SAMPLES, SO QUESTION 

918
00:32:18,328 --> 00:32:20,264
ABOUT CAN WE CREATE LOOK-ALIKES,

919
00:32:20,264 --> 00:32:21,598
MUCH LIKE THE EXAMPLES I SHOWED 

920
00:32:21,598 --> 00:32:23,834
WITH THE CT SCAN AND THAT'S 

921
00:32:23,834 --> 00:32:26,170
WHERE OUR THINKING IS RIGHT NOW.

922
00:32:26,170 --> 00:32:30,674
THE SECOND SUBSET THERE, DO THE 

923
00:32:30,674 --> 00:32:33,577
IMAGES -- CONVINCELY REAL 

924
00:32:33,577 --> 00:32:33,844
IMAGES?

925
00:32:33,844 --> 00:32:34,978
THIS IS A RECENT BLANK.

926
00:32:34,978 --> 00:32:36,980
WE DON'T HAVE AN ANSWER FOR 

927
00:32:36,980 --> 00:32:37,314
THIS.

928
00:32:37,314 --> 00:32:38,549
I WAS AT THE AMERICAN IMAGE 

929
00:32:38,549 --> 00:32:40,818
COMPUTING CONFERENCE JUST A WEEK

930
00:32:40,818 --> 00:32:42,519
AGO AND SOME GROUPS ARE TRYING 

931
00:32:42,519 --> 00:32:44,888
THIS IDEA WHERE THEY'RE HAVING 

932
00:32:44,888 --> 00:32:48,759
TWO CLASSIFIER PATHWAYS.

933
00:32:48,759 --> 00:32:50,961
ONE IS WHERE THE REAL IMAGE IS 

934
00:32:50,961 --> 00:32:52,796
MYLY I AM BALANCED, YOU BALANCE 

935
00:32:52,796 --> 00:32:55,666
IT OUT, USE TRADITIONAL 

936
00:32:55,666 --> 00:32:57,334
AUGMENTATION AND GET A 

937
00:32:57,334 --> 00:32:58,235
PERFORMANCE X.

938
00:32:58,235 --> 00:33:00,504
THE OTHER PATHWAY,OU RUN THE 

939
00:33:00,504 --> 00:33:02,539
LDM IN ONE CASE OR GANs, WHAT 

940
00:33:02,539 --> 00:33:04,508
HAVE YOU, AND YOU DON'T WORRY 

941
00:33:04,508 --> 00:33:06,844
ABOUT IT BEING A GOOD IMPRESSION

942
00:33:06,844 --> 00:33:07,644
OR NOT.

943
00:33:07,644 --> 00:33:10,247
YOU JUST -- YOU ASK FOR A LABEL,

944
00:33:10,247 --> 00:33:11,748
YOU GET WHAT IT IS AND THAT'S 

945
00:33:11,748 --> 00:33:12,783
THE LABEL YOU'RE GOING TO USE.

946
00:33:12,783 --> 00:33:16,353
SO NOW INCREASIN THE DIVER

947
00:33:16,353 --> 00:33:21,592
OF THE APPEARANCE WITH NO 

948
00:33:21,592 --> 00:33:23,093
VARIATION THAT THE SYNTHESIZED 

949
00:33:23,093 --> 00:33:24,428
IMAGES HAD ANYTHING TO DO WITH 

950
00:33:24,428 --> 00:33:27,497
THE DISEASE, BUT SOMEHOW, AN 

951
00:33:27,497 --> 00:33:28,699
EMSEM BELL OF THE TWO TEND BE TO

952
00:33:28,699 --> 00:33:31,001
BE BETTER THAN JUST USE A SMALL 

953
00:33:31,001 --> 00:33:31,401
SET.

954
00:33:31,401 --> 00:33:33,537
NOW, I NEED TO DWELL MORE ON 

955
00:33:33,537 --> 00:33:35,038
THIS, BUT THAT'S AN INTERESTING 

956
00:33:35,038 --> 00:33:35,272
THOUGHT.

957
00:33:35,272 --> 00:33:36,440
HERE SO FAR, WE'VE BEEN THINKING

958
00:33:36,440 --> 00:33:39,743
ABOUT SYNTHESIZE SOMETHING THAT 

959
00:33:39,743 --> 00:33:41,578
LOOKS LIKE IT, REPRESENTS THE 

960
00:33:41,578 --> 00:33:43,213
IMAGE THE WAY THE DISEASE AND 

961
00:33:43,213 --> 00:33:45,482
SEVERITY AS WE XN IT TO BE, SO 

962
00:33:45,482 --> 00:33:49,419
WE IN -- WE EXPECT IT TO BE SO 

963
00:33:49,419 --> 00:33:51,822
WE INCREASE THE SAMPLE SIZE 

964
00:33:51,822 --> 00:33:52,522
BELONGING TO THAT CLASS.

965
00:33:52,522 --> 00:33:54,725
IT'S A MIXTURE OF AUGMENTED AND 

966
00:33:54,725 --> 00:33:55,459
NON-AUGMENTED IMAGES.

967
00:33:55,459 --> 00:33:57,494
SO THAT IS AN OPEN THOUGHT AND I

968
00:33:57,494 --> 00:33:59,429
DON'T HAVE A GOOD ANSWER, AND 

969
00:33:59,429 --> 00:34:01,565
WE'RE LOOKING AT BOTH GENERATIVE

970
00:34:01,565 --> 00:34:05,502
AUGMENTATION AND TRADITIONAL 

971
00:34:05,502 --> 00:34:07,638
AUGMENTATION FOR NOT ONLY 

972
00:34:07,638 --> 00:34:08,538
TRADITIONAL X-RAYS, COMPARED 

973
00:34:08,538 --> 00:34:10,040
AGAINST NO CONDITION WHATSOEVER.

974
00:34:10,040 --> 00:34:11,041
THIS IS WORK WE'LL TALK ABOUT 

975
00:34:11,041 --> 00:34:12,876
NEXT, BUTNE THAT BULLET IS VERY 

976
00:34:12,876 --> 00:34:16,313
RECENT TALK THAT I'VE INSERTED 

977
00:34:16,313 --> 00:34:17,281
IN THERE BECAUSE OTHERS HAVE HIT

978
00:34:17,281 --> 00:34:20,784
THE SAME WALL THAT I'M GOING TO 

979
00:34:20,784 --> 00:34:22,786
SHOW YOU NOW AND ARE BYPASSING 

980
00:34:22,786 --> 00:34:24,888
IT SOMEWHAT IN WORK BY 

981
00:34:24,888 --> 00:34:26,490
JUST NOT CARING.

982
00:34:26,490 --> 00:34:27,925
AND I DON'T KNOW IF NOT CARING 

983
00:34:27,925 --> 00:34:30,327
IS A GOOD IDEA OR NOT.

984
00:34:30,327 --> 00:34:32,596
THEY'RE NOT DOING DISEASE 

985
00:34:32,596 --> 00:34:33,530
EXPLANATION OR HEAT MAPS BECAUSE

986
00:34:33,530 --> 00:34:34,398
THEY CAN'T.

987
00:34:34,398 --> 00:34:36,433
THEY DON'T HAVE THAT 

988
00:34:36,433 --> 00:34:36,733
INFORMATION.

989
00:34:36,733 --> 00:34:39,069
SO THAT IS DEFINITELY A 

990
00:34:39,069 --> 00:34:39,336
NEGATIVE.

991
00:34:39,336 --> 00:34:43,340
AND THE IDEA HERE IS WE'RE USING

992
00:34:43,340 --> 00:34:44,741
PROGRESSIVELY GROGAN WHICH 

993
00:34:44,741 --> 00:34:45,409
ESSENTIALLY SAMPLES THE IMAGES 

994
00:34:45,409 --> 00:34:49,112
AT DIFFERENT SIZES OF 

995
00:34:49,112 --> 00:34:52,015
RESOLUTION, AND YOU HAVE NO 

996
00:34:52,015 --> 00:34:53,116
OBLIGATION, YOU HAVE TRADITIONAL

997
00:34:53,116 --> 00:34:55,319
AUGMENTATION TECHNIQUES AS SHOWN

998
00:34:55,319 --> 00:34:57,521
IN THE BOTTOM ABOUT PATHWAY AND 

999
00:34:57,521 --> 00:34:59,122
GAN-PRODUCED IMAGES AND SEE IF 

1000
00:34:59,122 --> 00:35:00,891
THE CLASSIFIER PERFORMANCE 

1001
00:35:00,891 --> 00:35:01,925
ACTUALLY IMPROVES OR NOT.

1002
00:35:01,925 --> 00:35:05,095
IN THIS EXAMPLE, THE TOP RIGHT 

1003
00:35:05,095 --> 00:35:06,096
SHOWS HOW DIFFERENT RESOLUTION 

1004
00:35:06,096 --> 00:35:08,966
IMAGES ARE CREATED BYRE FG-GAN

1005
00:35:08,966 --> 00:35:11,368
AND YOU CAN FORCE THE PG-GAN TO 

1006
00:35:11,368 --> 00:35:15,739
 I WANT ONLY A FOURY FOUR 

1007
00:35:15,739 --> 00:35:18,342
IMAGE, AN EIGHT BY TEN IMAGE, SO

1008
00:35:18,342 --> 00:35:20,243
YOU SYNTHESIZE AN IMAGE THAT IS 

1009
00:35:20,243 --> 00:35:21,845
A LOOK-ALIKED CONDITIONED ON 

1010
00:35:21,845 --> 00:35:22,879
THAT RESOLUTION THAT YOU WANT.

1011
00:35:22,879 --> 00:35:23,347
YOU SEE W

1012
00:35:23,347 --> 00:35:23,947
ESOLUTION AND 

1013
00:35:23,947 --> 00:35:25,482
SCALE WAS AUGUST THE BEST 

1014
00:35:25,482 --> 00:35:26,416
AUGMENTATION TO THE END POINT AS

1015
00:35:26,416 --> 00:35:27,684
THE CLASSIFIER.

1016
00:35:27,684 --> 00:35:29,319
THAT'S WHAT WE'RE INTERESTED IN.

1017
00:35:29,319 --> 00:35:31,288
WE'RE NOT TRYING TO SYNTHESIZE 

1018
00:35:31,288 --> 00:35:32,422
PICTURES REALLY IN THIS EXAMPLE.

1019
00:35:32,422 --> 00:35:34,491
AND WE HAVE A GOOD NUMBER OF 

1020
00:35:34,491 --> 00:35:38,996
IMAGES, SO WE ACTUALLY USE THAT 

1021
00:35:38,996 --> 00:35:43,066
IN DIFFERENT RESOLUTIONS AND 

1022
00:35:43,066 --> 00:35:44,968
WHAT I'M GOING TO SHOW YOU HERE 

1023
00:35:44,968 --> 00:35:49,606
DID BETAT THE GANs. AUGMENTAT

1024
00:35:49,606 --> 00:35:51,808
SO REALLY, ALL THE EXTRA WORK 

1025
00:35:51,808 --> 00:35:54,277
PUT IN DID NOT MAKE AN IOTA OF 

1026
00:35:54,277 --> 00:35:55,212
DIFFERENCE IN THE PERFORMANCE.

1027
00:35:55,212 --> 00:35:58,382
AND THIS BEGS THE QUESTION, ARE 

1028
00:35:58,382 --> 00:35:58,849
WE -- BY

1029
00:35:58,849 --> 00:36:00,851
 ADDING MORE IMAGES, ARE

1030
00:36:00,851 --> 00:36:03,920
WE MOVING THE NEEDLE AT ALL?

1031
00:36:03,920 --> 00:36:06,590
AND IN THIS CASE, WE'VE TAKEN A 

1032
00:36:06,590 --> 00:36:07,924
WHACK A LITTLE.

1033
00:36:07,924 --> 00:36:12,362
HANG ON, WE DID ALL THIS WORK, 

1034
00:36:12,362 --> 00:36:13,730
THE PICTURES LOOK GOOD, AND YET 

1035
00:36:13,730 --> 00:36:16,233
WE'RE MAKING NO DIFFERENCE.

1036
00:36:16,233 --> 00:36:18,368
BECAUSE THE GANs ARE CREATING 

1037
00:36:18,368 --> 00:36:19,836
IMAGES THAT LOOK LIKE AND BELONG

1038
00:36:19,836 --> 00:36:21,605
TO THE SAME CLASS THAT THE 

1039
00:36:21,605 --> 00:36:24,641
DISCRIMINATOR WAS ASKED TO 

1040
00:36:24,641 --> 00:36:25,709
DISCRIMINATE AGAINST, THE 

1041
00:36:25,709 --> 00:36:27,077
FEATURE SPACE THAT'S OFF THE 

1042
00:36:27,077 --> 00:36:28,612
IMAGES THAT ARE BEING -- THEY'RE

1043
00:36:28,612 --> 00:36:30,881
USING FOR EXTRACTION BELONGS TO 

1044
00:36:30,881 --> 00:36:32,983
THE SAME CLUSTER, INCREASING THE

1045
00:36:32,983 --> 00:36:37,054
MORE TYPES OF, YOU'RE NOT 

1046
00:36:37,054 --> 00:36:38,355
SPREADING THE POINTS OUT, YOU'RE

1047
00:36:38,355 --> 00:36:41,024
NOT HAVING MORE DIVERSITY IN THE

1048
00:36:41,024 --> 00:36:44,261
PATTERNS, SO YOU'RE GETTING A 

1049
00:36:44,261 --> 00:36:48,131
PERHAPS MORE ELABORATE, BUT NOT 

1050
00:36:48,131 --> 00:36:51,034
 BETTER RESULTING -- THIS 

1051
00:36:51,034 --> 00:36:54,204
REPEATS AGAIN, AND SO THAT'S 

1052
00:36:54,204 --> 00:36:57,908
SOMETHING THAT WE WERE -- AND 

1053
00:36:57,908 --> 00:36:59,109
WORKING, AGAIN, THAT COMES TO 

1054
00:36:59,109 --> 00:37:00,410
MIND WITH THE OTHER COMMENT I 

1055
00:37:00,410 --> 00:37:01,678
MADE ABOUT SOME OTHER GROUPS 

1056
00:37:01,678 --> 00:37:04,548
TRYING TO BYPASS THIS WHOLE 

1057
00:37:04,548 --> 00:37:04,848
QUESTION.

1058
00:37:04,848 --> 00:37:06,917
BUT THEY'RE NOT DOING THE 

1059
00:37:06,917 --> 00:37:08,118
EXPLAINABILITY PART OF IT, SO 

1060
00:37:08,118 --> 00:37:08,952
THAT IS A COST.

1061
00:37:08,952 --> 00:37:10,554
THAT'S A PRICE YOU PAY.

1062
00:37:10,554 --> 00:37:11,922
MAYBE WORK FOR SOME 

1063
00:37:11,922 --> 00:37:12,255
APPLICATIONS.

1064
00:37:12,255 --> 00:37:14,124
SIMILAR WORK, IN THIS CASE 

1065
00:37:14,124 --> 00:37:17,227
WORKING WITH AI TO TRY TO 

1066
00:37:17,227 --> 00:37:19,496
DETERMINE IF ONE OR ANOTHER 

1067
00:37:19,496 --> 00:37:21,665
LESION IS PRECANCEROUS, AND TWO,

1068
00:37:21,665 --> 00:37:22,866
AFTER LESIONS ARE THERE, SHOULD 

1069
00:37:22,866 --> 00:37:25,102
IT BE BIOPSIED OR NOT.

1070
00:37:25,102 --> 00:37:28,738
THE CURRENT PROBLEM IS IT'S 100.

1071
00:37:28,738 --> 00:37:30,941
THE PATIENT NOTICES A LESION OR 

1072
00:37:30,941 --> 00:37:32,576
THE DENTIST NOTICES A LESION, 

1073
00:37:32,576 --> 00:37:35,245
SENDS YOU TO A CLING, THEY SEE A

1074
00:37:35,245 --> 00:37:38,181
LESION, THEY BIOPSY WHICH IS 

1075
00:37:38,181 --> 00:37:40,250
QUITE DISCOMFORT FOR PATIENTS 

1076
00:37:40,250 --> 00:37:41,485
ARE PRONE TO BE NEGATIVE AND FOR

1077
00:37:41,485 --> 00:37:43,086
THREE WEEKS, THEY'RE ON A 

1078
00:37:43,086 --> 00:37:44,321
DIFFERENT DIET ALSO ALTOGETHER 

1079
00:37:44,321 --> 00:37:45,489
AND SINGS LIKE THAT.

1080
00:37:45,489 --> 00:37:47,257
SO WE NEED -- AND THINGS LIKE 

1081
00:37:47,257 --> 00:37:47,457
THAT.

1082
00:37:47,457 --> 00:37:51,094
SO WE NEEDED TO HAVE MOVE 

1083
00:37:51,094 --> 00:37:54,164
EXAMPLES, SO WE HAVE AN 

1084
00:37:54,164 --> 00:37:56,666
IMBALANCE AGAIN WHERE THERE ARE 

1085
00:37:56,666 --> 00:37:59,769
CANCERS AND PRECANCERS AND WITHI

1086
00:37:59,769 --> 00:38:01,705
A NUMBER OF LESIONS, SO THE 

1087
00:38:01,705 --> 00:38:02,939
LESION NUMBER IS DIFFERENT FROM 

1088
00:38:02,939 --> 00:38:04,441
THE NUMBER OF PATIENTS THAT HAVE

1089
00:38:04,441 --> 00:38:04,841
CANCER.

1090
00:38:04,841 --> 00:38:11,114
AND IN THIS CASE, WE USED 

1091
00:38:11,114 --> 00:38:12,516
TEXT-IMAGE SYNTHESIZER WHERE YOU

1092
00:38:12,516 --> 00:38:14,618
HAVE THE MNL THAT CONTAINS THE 

1093
00:38:14,618 --> 00:38:16,419
EXISTING ORAL CANCER AND THE 

1094
00:38:16,419 --> 00:38:17,854
PROMPT IS ORAL A COLD FRONT WITH

1095
00:38:17,854 --> 00:38:20,490
CANCEROUS LESION AND THAT'S JUST

1096
00:38:20,490 --> 00:38:22,592
ADDING WORDS TO THE CLIP MODEL 

1097
00:38:22,592 --> 00:38:24,828
FOR IT TO LEARN FROM AND IT 

1098
00:38:24,828 --> 00:38:27,164
CONTAINS A CANCER TO PRECANCER 

1099
00:38:27,164 --> 00:38:29,199
SER RATIO OF NINE TO ONE AND YOU

1100
00:38:29,199 --> 00:38:31,535
GIVE IT THE IMAGES AND TRY TO 

1101
00:38:31,535 --> 00:38:32,903
SYNTHESIZE, AND THESE ARE SOME 

1102
00:38:32,903 --> 00:38:35,405
OF THE SYNTHESIZED IMAGES, 

1103
00:38:35,405 --> 00:38:35,672
ACTUALLY.

1104
00:38:35,672 --> 00:38:39,910
PICTURES.ATES PRETTYRE CONVINCI 

1105
00:38:39,910 --> 00:38:43,780
NOW, I THAT SOMEBODY COULD

1106
00:38:43,780 --> 00:38:47,884
COUNT THE TEETH AND SAY I 

1107
00:38:47,884 --> 00:38:49,619
DISAGREE WITH THE NUMBER OF 

1108
00:38:49,619 --> 00:38:52,322
TEETH, I DON'T SEE CANINE TEETH,

1109
00:38:52,322 --> 00:38:53,924
SO THAT SHOULD HAVE BEEN CATCH, 

1110
00:38:53,924 --> 00:38:54,925
BUT THE LESION PART IS PRETTY 

1111
00:38:54,925 --> 00:38:55,358
GOOD.

1112
00:38:55,358 --> 00:38:58,461
SO AGAIN, THE REGION OF 

1113
00:38:58,461 --> 00:38:59,396
INTEREST, IS THE IMAGE LOOKING 

1114
00:38:59,396 --> 00:39:00,163
ACCURATE OR NOT?

1115
00:39:00,163 --> 00:39:01,965
THAT'S SOMETHING TO THINK ABOUT.

1116
00:39:01,965 --> 00:39:04,267
THEN WE HAD DEVELOPED A SEPARATE

1117
00:39:04,267 --> 00:39:05,035
LESION SEGMENTATION MODEL THAT 

1118
00:39:05,035 --> 00:39:08,205
FOUND THE LESIONS FOR US, APPLY 

1119
00:39:08,205 --> 00:39:10,640
IT TO SYNTHESIZED IMAGES TO 

1120
00:39:10,640 --> 00:39:11,374
EXTRACT THE LESION.

1121
00:39:11,374 --> 00:39:13,043
SO AGAIN, IT'S FINDING THE 

1122
00:39:13,043 --> 00:39:14,945
LESIONS AND THE LESIONS ARE 

1123
00:39:14,945 --> 00:39:15,879
ACTUALLY REAL BECAUSE IT HAS 

1124
00:39:15,879 --> 00:39:17,113
BEEN TRAINED ON REAL IMAGES, SO 

1125
00:39:17,113 --> 00:39:20,750
IT FINDS THIS T AND WE CROP THE 

1126
00:39:20,750 --> 00:39:21,017
OUT.

1127
00:39:21,017 --> 00:39:23,486
SO YOU FOCUS ON AN IMAGE OF THE 

1128
00:39:23,486 --> 00:39:24,754
MOUTH, YOU HAVE LOTS OF LESION 

1129
00:39:24,754 --> 00:39:26,389
PATCHES NOW ON WHICH YOU WERE 

1130
00:39:26,389 --> 00:39:27,224
TRAINED.

1131
00:39:27,224 --> 00:39:29,559
AND IGNORE THE TEXT.

1132
00:39:29,559 --> 00:39:33,496
IT WILL MAKE A DIFFERENCE.

1133
00:39:33,496 --> 00:39:35,932
SO SOMETHING THAT IS NOW 

1134
00:39:35,932 --> 00:39:37,367
FOCUSED, REAL, ANOTHER 

1135
00:39:37,367 --> 00:39:38,335
CLASSIFIER BELIEVES THIS LOOKS 

1136
00:39:38,335 --> 00:39:41,705
LIKE A LESION THAT'S NOT RELATED

1137
00:39:41,705 --> 00:39:45,275
TO THE TRAINING CLASSIFIER AND 

1138
00:39:45,275 --> 00:39:48,211
WOE HIT THE SAME WALL AGAIN OF 

1139
00:39:48,211 --> 00:39:49,179
DIFFERENT NATURE, IMAGES WHICH 

1140
00:39:49,179 --> 00:39:52,582
ARE SHARP AND PRETTY LOOKING AND

1141
00:39:52,582 --> 00:39:53,650
FINE.

1142
00:39:53,650 --> 00:39:56,620
SO ARE WE ESSENTIALLY OBSERVING 

1143
00:39:56,620 --> 00:39:58,388
A DISTRIBUTION SHIFT HERE OR ARE

1144
00:39:58,388 --> 00:40:01,024
WE NOTICING THAT WE DON'T HAVE 

1145
00:40:01,024 --> 00:40:02,993
DIVERSITY DECISION AS I SAID 

1146
00:40:02,993 --> 00:40:03,627
EARLIER.

1147
00:40:03,627 --> 00:40:03,994
I DON'T KNOW.

1148
00:40:03,994 --> 00:40:05,061
THIS IS QUESTIONS THAT -- OF 

1149
00:40:05,061 --> 00:40:06,496
SOMETHING TO QUAUN DER ABOUT AND

1150
00:40:06,496 --> 00:40:10,567
WE DON'T HAVE GOOD ANSWERS FOR.

1151
00:40:10,567 --> 00:40:11,868
ON THE FLIP SIDE, WE'RE GOING TO

1152
00:40:11,868 --> 00:40:15,338
HAVE SOME GOOD NEWS NOW, WHICH 

1153
00:40:15,338 --> 00:40:19,242
IS USING LATON MODELS

1154
00:40:19,242 --> 00:40:20,210
IN PEDIATRIC CHEST IN, WHICH

1155
00:40:20,210 --> 00:40:22,779
IS WHAT I WAS ALLUDING TO 

1156
00:40:22,779 --> 00:40:23,513
EARLIER.

1157
00:40:23,513 --> 00:40:25,915
AGAIN, THERE'S THIS CLASS 

1158
00:40:25,915 --> 00:40:27,117
IMBALANCE WHERE YOU HAVE FAR 

1159
00:40:27,117 --> 00:40:30,720
MORE -- FAR MORE NORMALS THAN 

1160
00:40:30,720 --> 00:40:33,823
DISEASE POSITIVE, WHICH IS, AS A

1161
00:40:33,823 --> 00:40:35,759
WORD, WE'RE HEALTHY.

1162
00:40:35,759 --> 00:40:36,226
WE'RE

1163
00:40:36,226 --> 00:40:39,162
 SAMPLING THINGS WITH FAR 

1164
00:40:39,162 --> 00:40:41,264
FEWER PEOPLE WHO ARE EXPOSED TO 

1165
00:40:41,264 --> 00:40:44,734
RADIATION FOR THE FUN OF IT.

1166
00:40:44,734 --> 00:40:50,340
AND THE ARGUMENT HERE, IS 

1167
00:40:50,340 --> 00:40:50,974
TRADITIONAL AUGMENT SUFFICIENT 

1168
00:40:50,974 --> 00:40:52,309
AND IS IT SUFFICIENT MAINLY 

1169
00:40:52,309 --> 00:40:55,111
BECAUSE OF THE POSE OF THE KID 

1170
00:40:55,111 --> 00:40:56,913
WHO IS IMAGED.

1171
00:40:56,913 --> 00:40:59,816
BECAUSE IF YOU HAVE -- FOR ADULT

1172
00:40:59,816 --> 00:41:02,185
X-RAY, IT'S A LOT MORE 

1173
00:41:02,185 --> 00:41:04,688
STRAIGHTFORWARD, BUT THE PERSON 

1174
00:41:04,688 --> 00:41:06,890
IS AGAINST PLATE, THEY'RE 

1175
00:41:06,890 --> 00:41:12,329
FRAMED WELL, ENSURED BY THE 

1176
00:41:12,329 --> 00:41:14,264
RADIOING ON  -- RADIOGRAPHER AND

1177
00:41:14,264 --> 00:41:16,032
YOU REPEATABLE IMAGES.

1178
00:41:16,032 --> 00:41:19,002
KIDS ALL OVER THE PLACE, AGE 

1179
00:41:19,002 --> 00:41:21,338
MATTERS A LOT.

1180
00:41:21,338 --> 00:41:22,972
LUNG SIZE IS DIFFERENT.

1181
00:41:22,972 --> 00:41:25,742
KIDS BELOW 5 YEARS OF AGE, 

1182
00:41:25,742 --> 00:41:27,644
SMALLER, BOYS ARE LATER IN TAT T

1183
00:41:27,644 --> 00:41:28,812
LIVES BY, LOOKING MORE 

1184
00:41:28,812 --> 00:41:30,347
LIKE ADULTS, SO THERE'S THAT 

1185
00:41:30,347 --> 00:41:32,349
TRAIN THAT IS THERE IN THE SHAPE

1186
00:41:32,349 --> 00:41:33,483
OF THE OBJECT OF INTEREST AS 

1187
00:41:33,483 --> 00:41:37,721
WELL, REGION OF INTEREST.

1188
00:41:37,721 --> 00:41:40,757
AND WE HAVE 9,000 IMAGES WE CAN 

1189
00:41:40,757 --> 00:41:41,624
PULL PUBLICLY FROM REFERENCE 

1190
00:41:41,624 --> 00:41:43,093
NUMBER ONE AT THE BOTTOM HERE 

1191
00:41:43,093 --> 00:41:46,996
AND THEY HAVE TWO KINDS, 

1192
00:41:46,996 --> 00:41:50,800
PNEUMONIA AND WE DO THE CLASSIC 

1193
00:41:50,800 --> 00:41:51,801
COMPARISON.

1194
00:41:51,801 --> 00:41:54,304
THIS IS PNEUMONIA, BRONGO 

1195
00:41:54,304 --> 00:41:56,639
PNEUMONIA AND THERE'S A 20% FOR 

1196
00:41:56,639 --> 00:41:57,941
HOLD-OUT TESTING, THE CLASSIC 

1197
00:41:57,941 --> 00:42:00,076
BREAK YOU KNOW.

1198
00:42:00,076 --> 00:42:05,248
AND WE USE ULV5 TO FIND THE 

1199
00:42:05,248 --> 00:42:08,651
REGION OF INTEREST, SO NORTH 

1200
00:42:08,651 --> 00:42:09,953
WHAT THE POSE IS -- NO MATTER 

1201
00:42:09,953 --> 00:42:12,021
WHAT THE POSE IS, WE LOCK DOWN 

1202
00:42:12,021 --> 00:42:13,590
ON THE REGION OF INTEREST AND 

1203
00:42:13,590 --> 00:42:15,225
WHAT WE WANT TO FOCUS ON.

1204
00:42:15,225 --> 00:42:16,826
IT LOOKS LIKE A LOT LIKE THE 

1205
00:42:16,826 --> 00:42:18,061
IMAGE YOU'RE SEEING IN THE 

1206
00:42:18,061 --> 00:42:19,662
EXAMPLE ON TOP, KIDS OF 

1207
00:42:19,662 --> 00:42:21,931
DIFFERENT AGES THERE.

1208
00:42:21,931 --> 00:42:24,801
NOW YOU HAVE A DECISION WITH AN 

1209
00:42:24,801 --> 00:42:28,071
INCIDENCE THAT SAYS CHEST X-RAY 

1210
00:42:28,071 --> 00:42:32,742
SHOWING PNEUMONIA AND IT -- 

1211
00:42:32,742 --> 00:42:34,844
RESOLUTION GOING ON AND YOU 

1212
00:42:34,844 --> 00:42:36,045
HAVED PEDIATRIC X-RAY 

1213
00:42:36,045 --> 00:42:40,016
WHICH IS NORMAL, AND IT LEARNS 

1214
00:42:40,016 --> 00:42:45,488
TOTO SYNTHESIZE IMAGES THAT ARE 

1215
00:42:45,488 --> 00:42:47,223
WHAT YOU WANT AND IT IS GOING TO

1216
00:42:47,223 --> 00:42:48,525
BE GENERATING IMAGES WITH A 

1217
00:42:48,525 --> 00:42:50,527
PIECE OF TEXT AND YOUNT TO 

1218
00:42:50,527 --> 00:42:52,195
SYNTHESIZE IMAGES THAT LOOK LIKE

1219
00:42:52,195 --> 00:42:53,930
THAT IMAGE.

1220
00:42:53,930 --> 00:42:56,533
IN YOUR ASK, YOU'RE NOT USING A 

1221
00:42:56,533 --> 00:42:58,501
TEXT PROMPT TO GIVE YOU IMAGES, 

1222
00:42:58,501 --> 00:43:00,270
YOU'RE USING TEXT PLUS PROMPT.

1223
00:43:00,270 --> 00:43:02,705
HERE'S AN EXAMPLE IMAGE, WE NEED

1224
00:43:02,705 --> 00:43:05,074
ALL SAMPLES THAT BELONG THO THIS

1225
00:43:05,074 --> 00:43:06,042
IMAGE TYPE.

1226
00:43:06,042 --> 00:43:08,244
THERE'S A BIT OF A SHIFT GOING 

1227
00:43:08,244 --> 00:43:11,981
ON HERE FROM THE GAN-BASED 

1228
00:43:11,981 --> 00:43:13,583
APPROACHES WHERE THEY'RE PUTTING

1229
00:43:13,583 --> 00:43:15,852
THE DECISION MODEL INTO -- 

1230
00:43:15,852 --> 00:43:18,555
TAKING ADVANTAGE OF THE VARIETY 

1231
00:43:18,555 --> 00:43:20,323
THAT EXISTS.

1232
00:43:20,323 --> 00:43:21,458
WHEN YOU HAVE THOUSANDS OF 

1233
00:43:21,458 --> 00:43:22,492
X-RAYS TO BEGIN WITH.

1234
00:43:22,492 --> 00:43:24,461
SO THERE'S SOME BENEFIT TO THIS 

1235
00:43:24,461 --> 00:43:25,495
APPROACH THAT'S DIFFERENT FROM 

1236
00:43:25,495 --> 00:43:27,997
THE PAST APPROACHES.

1237
00:43:27,997 --> 00:43:30,667
AND YOU NOW HAVE THE EXAMPLE OF 

1238
00:43:30,667 --> 00:43:31,434
PEDIATRIC CHEST X-RAYS SHOWING 

1239
00:43:31,434 --> 00:43:33,536
KNEW MOAHO THE NORMAL, AND

1240
00:43:33,536 --> 00:43:34,804
THE SYNTHESIZED, HOW IT 

1241
00:43:34,804 --> 00:43:38,441
TRANSLATES ONE TO THE OTHER, AND

1242
00:43:38,441 --> 00:43:43,012
YOU HAVE BRONCHO PNEUMONIA GOING

1243
00:43:43,012 --> 00:43:44,214
FROM ONE TO THE OTHER.

1244
00:43:44,214 --> 00:43:46,282
SO VISUALLY, IT IS DOING 

1245
00:43:46,282 --> 00:43:47,116
BELIEVABLE THINGS WHEN YOU LOOK 

1246
00:43:47,116 --> 00:43:48,551
AT NUMBER OF SAMPLES, GOING FROM

1247
00:43:48,551 --> 00:43:49,752
NORMAL TO ABNORMAL, AND 

1248
00:43:49,752 --> 00:43:51,054
THEREFORE INCREASING THE NUMBER 

1249
00:43:51,054 --> 00:43:52,789
OF ABNORMAL THAT WE HAVE IN THE 

1250
00:43:52,789 --> 00:43:54,524
SET SO THAT THE CLASSIFIER 

1251
00:43:54,524 --> 00:43:55,992
THAT'S GOING TO BE TRAINED ON 

1252
00:43:55,992 --> 00:43:58,328
 IMAGES IS SEEING A BALANCE

1253
00:43:58,328 --> 00:44:02,365
ON THE OTHER SIDE.

1254
00:44:02,365 --> 00:44:06,169
SO YOU HAVE AN IMAGE THAT IS 

1255
00:44:06,169 --> 00:44:07,704
TRAINED BETWEEN TWO MODELS, 

1256
00:44:07,704 --> 00:44:10,240
TRAINED ON A CLASS IMBALANCE 

1257
00:44:10,240 --> 00:44:13,209
TEST FIRST, THAT'S YOUR BASELINE

1258
00:44:13,209 --> 00:44:16,980
FOR NORMAL VERSUS PNEUMONIA 

1259
00:44:16,980 --> 00:44:18,848
YOU HAVE ADD TO ITHA TO IT

1260
00:44:18,848 --> 00:44:20,717
OUT AND YOU HAVE A NEW SET THAT 

1261
00:44:20,717 --> 00:44:23,753
INCLUDES A MIX OF SYNTHETIC AND 

1262
00:44:23,753 --> 00:44:25,221
NONSYNTHETIC IMAGES, BUT YOU'RE 

1263
00:44:25,221 --> 00:44:26,823
TESTING ONLY ON THE SYNTHETIC, 

1264
00:44:26,823 --> 00:44:28,758
ON THE -- I BEG YOUR PARDON, ON 

1265
00:44:28,758 --> 00:44:29,659
THE REAL IMAGES.

1266
00:44:29,659 --> 00:44:31,094
YOU CAN'T TEST ON SYNTHETIC, YOU

1267
00:44:31,094 --> 00:44:32,495
DON'T KNOW WHAT IT IS.

1268
00:44:32,495 --> 00:44:35,098
AND YOU TEST FOR A VARIETY OF 

1269
00:44:35,098 --> 00:44:35,331
METRICS.

1270
00:44:35,331 --> 00:44:36,666
THE IMPORTANT PART, IF Y  F

1271
00:44:36,666 --> 00:44:38,501
TO TAKE IT INTO THE FIELD, YOU 

1272
00:44:38,501 --> 00:44:40,403
USE THE INDEX WHICH TELLS --

1273
00:44:40,403 --> 00:44:41,771
YOU KNOW, WHAT IS THE BALANCE OF

1274
00:44:41,771 --> 00:44:43,706
INTENSITY AND SPECIFICITY OF A 

1275
00:44:43,706 --> 00:44:45,742
TEST AS IT GOES INTO THE FIELD 

1276
00:44:45,742 --> 00:44:48,778
AND YOU SEE IT'S A LOT -- THE 

1277
00:44:48,778 --> 00:44:50,880
COEFFICIENT IN TERMS OF 

1278
00:44:50,880 --> 00:44:52,482
COMPARING HOW MANY OF THEM ALONG

1279
00:44:52,482 --> 00:44:55,785
THE DIAGONAL AND HOW MUCH 

1280
00:44:55,785 --> 00:44:56,553
AGREEMENT THERE IS.

1281
00:44:56,553 --> 00:44:59,556
THE COEFFICIENT IS USING MACHINE

1282
00:44:59,556 --> 00:45:00,657
LEARNING, COEFFICIENT IN A 

1283
00:45:00,657 --> 00:45:04,661
THE PEARSON'S TEST VERSION OF NN

1284
00:45:04,661 --> 00:45:05,828
IT, SO DIFFERENT NAMES FOR THE 

1285
00:45:05,828 --> 00:45:09,165
SAME KIND OF TEST BASICALLY.

1286
00:45:09,165 --> 00:45:10,166
AND YOU'RE SEEING SLIGHTLY 

1287
00:45:10,166 --> 00:45:12,335
BETTER RESULTS FOR THE AUGMENTED

1288
00:45:12,335 --> 00:45:12,802
ONES.

1289
00:45:12,802 --> 00:45:15,872
BASELINE VERSUS AUGMENTED, 

1290
00:45:15,872 --> 00:45:18,808
SPECIFICITY WISE, BASELINE STILL

1291
00:45:18,808 --> 00:45:20,710
BETTER AND MAYBE THE NORMALS ARE

1292
00:45:20,710 --> 00:45:23,112
DON'T KNOW.PECIAL ABOUT THEM,

1293
00:45:23,112 --> 00:45:26,316
WE HAVEN'T VERIFIED THIS IN 

1294
00:45:26,316 --> 00:45:32,989
TERMS OF HOW WHY IS THE 

1295
00:45:32,989 --> 00:45:33,690
SPECIFICITY TEST IMPROVING AT 

1296
00:45:33,690 --> 00:45:35,992
THE COST OF A FEWSAMPLES, AND 

1297
00:45:35,992 --> 00:45:37,160
YOU SEE THE TWO TABLES BELOW 

1298
00:45:37,160 --> 00:45:39,128
WHERE YOU SEE HERE, THE 

1299
00:45:39,128 --> 00:45:41,631
AUGMENTED CLASS WENT DOWN BY 7 

1300
00:45:41,631 --> 00:45:46,102
IN THIS CASE AND -- BUT YOU 

1301
00:45:46,102 --> 00:45:48,571
GAINED MORE TRUE POSITIVES AT 

1302
00:45:48,571 --> 00:45:50,039
THE COST OF SOME FALSE NEGATIVES

1303
00:45:50,039 --> 00:45:51,674
AS WELL -- FALSE POSITIVES AS 

1304
00:45:51,674 --> 00:45:51,874
WELL.

1305
00:45:51,874 --> 00:45:53,610
SO THE FALSE NEGATIVES ARE 

1306
00:45:53,610 --> 00:45:56,346
NOT -- SO THAT IS THE PRICE 

1307
00:45:56,346 --> 00:45:58,448
PAYING.

1308
00:45:58,448 --> 00:46:02,352
ON THE WHOLE, THIS IS TRAINING 

1309
00:46:02,352 --> 00:46:04,721
TOWARD AT LEAST CONFIRMING YOUR 

1310
00:46:04,721 --> 00:46:08,825
HYPOTHESIS THAT GAN-BASED 

1311
00:46:08,825 --> 00:46:10,660
REPLICATION IS GOOD FOR CHANGING

1312
00:46:10,660 --> 00:46:13,963
THE APPEARANCE OF AN IMAGE, BUT 

1313
00:46:13,963 --> 00:46:16,566
NOTSARILYTT ADDING 

1314
00:46:16,566 --> 00:46:19,068
STATISTICAL VALUE TO THE FEATURE

1315
00:46:19,068 --> 00:46:19,569
DISTRIBUTION.

1316
00:46:19,569 --> 00:46:20,770
LDM-BASED DOES BETTER.

1317
00:46:20,770 --> 00:46:23,106
NOW, THIS PARTICULAR SET, WHICH 

1318
00:46:23,106 --> 00:46:27,510
IS A GIGANTIC CONDITION HERE, 

1319
00:46:27,510 --> 00:46:31,714
HELPED POINT US IN THE RIGHT 

1320
00:46:31,714 --> 00:46:32,715
DIRECTION.

1321
00:46:32,715 --> 00:46:34,250
CAN WE ACTUALLY STILL USE HEAT 

1322
00:46:34,250 --> 00:46:36,719
MAPS TO TELL WHERE YOU'RE 

1323
00:46:36,719 --> 00:46:38,688
LOOKING, FOR KIDS ABOVE THIS -- 

1324
00:46:38,688 --> 00:46:40,023
THE REASON THEY'RE WORKING ON 

1325
00:46:40,023 --> 00:46:41,958
THIS IS WE ARE HEADING TOWARD A 

1326
00:46:41,958 --> 00:46:49,432
CHEST X-RAYS FOR KIDS WITHOUT 

1327
00:46:49,432 --> 00:46:54,971
HI HAVE TB BECAUSE THEY 

1328
00:46:54,971 --> 00:47:01,778
DON'T -- EXPECTORATE, SO IT'S 

1329
00:47:01,778 --> 00:47:03,813
OFTEN MISDIAGNOSED AND IT'S A 

1330
00:47:03,813 --> 00:47:05,248
MORTALITY FOR KIDS WHO ARE 

1331
00:47:05,248 --> 00:47:05,915
IMMUNE COMPROMISED.

1332
00:47:05,915 --> 00:47:10,920
SO WHAT YOU WANT TO DO IS TELL 

1333
00:47:10,920 --> 00:47:14,457
YOU IN MOAN YEAH FROM TB.

1334
00:47:14,457 --> 00:47:17,660
-- KNEW MONA FROM TB.

1335
00:47:17,660 --> 00:47:19,362
YOU CAN TRAIN AN OBSERVER WITHIN

1336
00:47:19,362 --> 00:47:20,830
A DAY TO LOOK AT THESE PATTERNS,

1337
00:47:20,830 --> 00:47:22,532
BUT FOR KIDS, THESE PATTERNS 

1338
00:47:22,532 --> 00:47:23,666
DON'T EXIST.

1339
00:47:23,666 --> 00:47:24,333
THEIR LYMPH NODES EXPAND AND 

1340
00:47:24,333 --> 00:47:25,702
THEY GET OTHER EXPRESSIONS ON 

1341
00:47:25,702 --> 00:47:30,139
THE LUNG THAT LOOKS A LOT LIKE 

1342
00:47:30,139 --> 00:47:32,041
PNEUMONIA, BUT IS NOT, IT'S TB, 

1343
00:47:32,041 --> 00:47:33,376
AND THERE'S A MAGICAL 

1344
00:47:33,376 --> 00:47:35,845
COMBINATION OF THIS THAT IS 

1345
00:47:35,845 --> 00:47:37,947
EMBEDDED IN THE DISEASE THAT 

1346
00:47:37,947 --> 00:47:40,750
WE'RE TRYING TO EXTRAPOLATE OUT 

1347
00:47:40,750 --> 00:47:41,984
BY THEIR LABELS AND WE GOT LABEL

1348
00:47:41,984 --> 00:47:45,188
JUST COMPLETING ON A SET OF A 

1349
00:47:45,188 --> 00:47:47,523
THOUSAND IMAGES OF 200 OF WHICH 

1350
00:47:47,523 --> 00:47:49,892
ARE NORMALS AND THE AEMAINING 

1351
00:47:49,892 --> 00:47:53,563
800 AND CHANGE ARE ABNORMALS.

1352
00:47:53,563 --> 00:47:54,964
THE CHALLENGE IS GOING TO 

1353
00:47:54,964 --> 00:47:56,499
PRESENT ITSELF BECAUPR THESE ARE

1354
00:47:56,499 --> 00:47:57,834
SELECTED FROM 15 SITES 

1355
00:47:57,834 --> 00:48:00,103
WORLDWIDE, SO WE DON'T GET THAT 

1356
00:48:00,103 --> 00:48:01,270
EXPOSURE VARIATION THAT 

1357
00:48:01,270 --> 00:48:02,872
HAPPENS -- WE GET THAT EXPOSURE 

1358
00:48:02,872 --> 00:48:04,607
VARIATION THAT HAPPENS BECAUSE 

1359
00:48:04,607 --> 00:48:05,475
OF THE PROTOCOL AT DIFFERENT 

1360
00:48:05,475 --> 00:48:06,309
CLINICS AND WE HAVE TO 

1361
00:48:06,309 --> 00:48:08,978
COMPENSATE FOR THAT, SO MAYBE 

1362
00:48:08,978 --> 00:48:10,446
THERE WILL BE SOME GANs PLAYING 

1363
00:48:10,446 --> 00:48:13,282
A ROLE, NOT SURE YET, BUT THAT'S

1364
00:48:13,282 --> 00:48:15,685
WHAT WE ARE -- WE KNOW WE'RE ON 

1365
00:48:15,685 --> 00:48:17,220
THAT TREADMILL HEADING TOWARD 

1366
00:48:17,220 --> 00:48:19,322
THAT, AND SO THIS IS ALL 

1367
00:48:19,322 --> 00:48:20,957
BUILDING OUR BACK KNOWLEDGE FOR 

1368
00:48:20,957 --> 00:48:22,792
ANSWERING THE QUESTION, AND IT'S

1369
00:48:22,792 --> 00:48:25,695
HEARTENING TO KNOW THAT LDM 

1370
00:48:25,695 --> 00:48:28,731
ACTUALLY IS BETTER THAN GANs IN 

1371
00:48:28,731 --> 00:48:30,032
BALANCING THE SLA ANDND

1372
00:48:30,032 --> 00:48:32,435
PRODUCING HIGH QUALITY IMAGESGE 

1373
00:48:32,435 --> 00:48:34,303
AND GIVING CLASSIFIER WORKS THAT

1374
00:48:34,303 --> 00:48:35,204
ARE SOMEWHAT WERN THE 

1375
00:48:35,204 --> 00:48:36,572
PREVIOUS ONE, SIGNIFICANTLY 

1376
00:48:36,572 --> 00:48:38,107
BETTER IN THE INDEX FOR SURE, 

1377
00:48:38,107 --> 00:48:41,911
BUT NOT ENOUGH TO GO OUT TO THE 

1378
00:48:41,911 --> 00:48:44,213
FIELD BECAUSE THE INDEX IS NOT 

1379
00:48:44,213 --> 00:48:48,117
WORTH LOOKING AT FROM A FIELD 

1380
00:48:48,117 --> 00:48:50,019
TESTING POINT OF VIEW.

1381
00:48:50,019 --> 00:48:52,455
AND THAT'S BECAUSE IT IS HEAVILY

1382
00:48:52,455 --> 00:48:53,956
SKEWED TOWARD SPECIFICITY AS IS 

1383
00:48:53,956 --> 00:48:54,690
OBJECT IN THE SCORE.

1384
00:48:54,690 --> 00:48:56,325
SO THE MODEL IS ONLY AS GOOD AS 

1385
00:48:56,325 --> 00:48:57,460
THE DATA AND THE PROBLEM AT 

1386
00:48:57,460 --> 00:48:58,895
HAND, WHICH IS NOT MENTIONED 

1387
00:48:58,895 --> 00:49:01,164
HERE, BUT IS IMPLIED, AND 

1388
00:49:01,164 --> 00:49:04,567
THEREFORE, I PROPERLY SHOWED YOU

1389
00:49:04,567 --> 00:49:08,271
THAT THE IMAGE SYNTHESIS USING 

1390
00:49:08,271 --> 00:49:11,340
GANs AND LDMs ARE A VIABLE 

1391
00:49:11,340 --> 00:49:12,308
APPROACH, BUT CHALLENGING.

1392
00:49:12,308 --> 00:49:13,376
YOU HAVE TO KNOW YOUR PROBLEM, 

1393
00:49:13,376 --> 00:49:14,777
KNOW THE DATA, KNOW THE 

1394
00:49:14,777 --> 00:49:16,012
DISTRIBUTION AND THE QUALITY OF 

1395
00:49:16,012 --> 00:49:17,713
THE IMAGES AND THE KINDS OF 

1396
00:49:17,713 --> 00:49:19,415
THINGS YOU'RE LOOKING FOR BEFORE

1397
00:49:19,415 --> 00:49:20,583
YOU CAN START USING A PARTICULAR

1398
00:49:20,583 --> 00:49:22,585
APPROACH OR A COMBINATION OF 

1399
00:49:22,585 --> 00:49:24,220
APPROACHES.

1400
00:49:24,220 --> 00:49:25,688
CONTEXT IS KEY, AND CONTINUOUS 

1401
00:49:25,688 --> 00:49:26,589
EVALUATION AT DIFFERENT STAGES 

1402
00:49:26,589 --> 00:49:29,158
IS CRITICAL.

1403
00:49:29,158 --> 00:49:32,028
WITH THAT, I I THANK YOU I Y FO 

1404
00:49:32,028 --> 00:49:34,363
ATTENTION AND IF YOU HAVE ANY 

1405
00:49:34,363 --> 00:49:34,630
QUESTIONS.

1406
00:49:34,630 --> 00:49:43,005
[ APPLAUSE ]

1407
00:49:43,005 --> 00:49:46,242
>> SO THANK YOU FOR THE
TALK.

1408
00:49:46,242 --> 00:49:48,377
SO BEFORE WE START THE Q&A 

1409
00:49:48,377 --> 00:49:51,948
SESSION HERE, FOR THE PEOPLE WHO

1410
00:49:51,948 --> 00:49:53,883
ARE WATCHING US ON -- FOR THE 

1411
00:49:53,883 --> 00:49:55,551
PEOPLE WATCHING US ON VIDEO 

1412
00:49:55,551 --> 00:49:56,686
CAST, PLEASE SEND YOUR QUESTIONS

1413
00:49:56,686 --> 00:50:00,223
VIA -- THERE'S A BUTTON LABELED 

1414
00:50:00,223 --> 00:50:03,292
LIVE FEEDBACK, AND YOU CAN SEND 

1415
00:50:03,292 --> 00:50:04,327
US YOUR QUESTIONS THROUGH THERE.

1416
00:50:04,327 --> 00:50:05,862
I'LL KEEP AN EYE ON THOSE 

1417
00:50:05,862 --> 00:50:06,128
QUESTIONS.

1418
00:50:06,128 --> 00:50:10,366
SO FOR THE PEOPLE HERE IN THE 

1419
00:50:10,366 --> 00:50:13,970
ROOM, ANY QUESTIONS?

1420
00:50:13,970 --> 00:50:19,609
LET ME HAND YOU THE MICROPHONE.

1421
00:50:19,609 --> 00:50:23,112
>> I'M NOT FAMILIAR WITH
THE 

1422
00:50:23,112 --> 00:50:26,883
MODEL THAT YOU MENTIONED, 

1423
00:50:26,883 --> 00:50:30,987
DEBLURRING -- DEBLUR GAN.

1424
00:50:30,987 --> 00:50:32,922
IN A CLASSICAL GAN, YOU TRAIN 

1425
00:50:32,922 --> 00:50:36,993
THE MODEL ON GOOD DATA AND THEN 

1426
00:50:36,993 --> 00:50:40,830
TRY TO GENERATE AS GOOD AS THE 

1427
00:50:40,830 --> 00:50:42,365
DATA IMAGES FROM SCRATCH, FROM 

1428
00:50:42,365 --> 00:50:43,900
NOISE.

1429
00:50:43,900 --> 00:50:48,838
IN THE BLUR GAN, I'M ASSUMING --

1430
00:50:48,838 --> 00:50:51,574
DEBLUR GAN, I ASSUME INSTEAD OF 

1431
00:50:51,574 --> 00:50:52,541
NOISE, YOU'RE USING A REAL 

1432
00:50:52,541 --> 00:50:53,042
IMAGE. 

1433
00:50:53,042 --> 00:50:54,343
>> A PAIR OF IMAGES.

1434
00:50:54,343 --> 00:50:56,145
ONE THAT IS THE GOOD IMAGE AND 

1435
00:50:56,145 --> 00:50:59,715
ONE THAT IS THE NOISY IMAGE.

1436
00:50:59,715 --> 00:51:01,851
THAT'S GOING TO DISCRIMINATOR TO

1437
00:51:01,851 --> 00:51:03,286
TELL APART ONE FROM THE OTHER 

1438
00:51:03,286 --> 00:51:05,054
AND THE BLURRY IMAGE IS BEING 

1439
00:51:05,054 --> 00:51:07,990
FED TO THE GENERATOR. 

1440
00:51:07,990 --> 00:51:10,192
>> YOU TRY TO MOVE FROM
BLURRED 

1441
00:51:10,192 --> 00:51:11,093
TO IDEAL IMAGE. 

1442
00:51:11,093 --> 00:51:11,427
>> CORRECT. 

1443
00:51:11,427 --> 00:51:13,896
>> AND KIND OF STOP AT
SOME 

1444
00:51:13,896 --> 00:51:15,431
POINT WHERE -- 

1445
00:51:15,431 --> 00:51:17,133
>> WHEN THE LAST FUNCTION,
WHICH

1446
00:51:17,133 --> 00:51:20,069
TRIES TO -- WHICH TRIES TO 

1447
00:51:20,069 --> 00:51:22,371
MINIMIZE THE GAP BETWEEN 

1448
00:51:22,371 --> 00:51:25,408
BLURRY -- I'M SORRY, A DEBLURED 

1449
00:51:25,408 --> 00:51:27,143
IMAGE AND THE AVERAGE DEBLURRED 

1450
00:51:27,143 --> 00:51:28,077
IMAGE, IF YOU WILL.

1451
00:51:28,077 --> 00:51:29,979
>> OKAY. 

1452
00:51:29,979 --> 00:51:31,681
>> AND SO IT IS USING A 

1453
00:51:31,681 --> 00:51:33,215
DIFFERENT LOSS FUNCTION FOR THE 

1454
00:51:33,215 --> 00:51:35,151
SCRIM NATOR AND A DIFFERENT 

1455
00:51:35,151 --> 00:51:37,186
FUNCTION FOR THE GENERATOR, A 

1456
00:51:37,186 --> 00:51:37,553
CONDITIONAL GAN.

1457
00:51:37,553 --> 00:51:38,321
IT'S CONDITIONED ON THE IDEA 

1458
00:51:38,321 --> 00:51:40,323
THAT YOU HAVE GLUR RI AND 

1459
00:51:40,323 --> 00:51:41,857
NONBLURRY IMAGES.

1460
00:51:41,857 --> 00:51:43,759
SO IT IS A DIFFERENT CONSTRUCT 

1461
00:51:43,759 --> 00:51:46,395
FOR -- BECAUSE IF YOU'RE TRYING 

1462
00:51:46,395 --> 00:51:50,066
TO TRANSLATE A BLURRY IMAGE INTO

1463
00:51:50,066 --> 00:51:51,067
A NONEXISTENT DEBLURRED IMAGE 

1464
00:51:51,067 --> 00:51:52,969
FROM THE REAL WORLD, SO WE DON'T

1465
00:51:52,969 --> 00:51:54,103
WANT TO CHANGE THE IMAGE.

1466
00:51:54,103 --> 00:51:56,539
YOU WANT TO KEEP THE SAME 

1467
00:51:56,539 --> 00:51:58,374
CONTENT OR CONTEXT AS IT MIGHT 

1468
00:51:58,374 --> 00:52:00,776
HAVE BEEN HAD IT BEEN DEBLURRED 

1469
00:52:00,776 --> 00:52:01,911
AND THAT'S WHY YOU HAVE THE 

1470
00:52:01,911 --> 00:52:03,813
NOISE RATIO COMPARISON BECAUSE 

1471
00:52:03,813 --> 00:52:05,581
WE ACTUALLY BLURRED THE IMAGES 

1472
00:52:05,581 --> 00:52:07,984
OURSELVES, WE KNOW WHAT THAT IS,

1473
00:52:07,984 --> 00:52:10,619
SO HOW FAR IS THE DEBLURRED 

1474
00:52:10,619 --> 00:52:14,190
VERSION FROM THE ORIGINAL.

1475
00:52:14,190 --> 00:52:15,591
THAT'S ONE THING ON THE VALUE 

1476
00:52:15,591 --> 00:52:17,493
AND THE OTHER IS THE INDEX WHICH

1477
00:52:17,493 --> 00:52:18,794
TELLS YOU IF YOU WERE TO 

1478
00:52:18,794 --> 00:52:20,796
ESSENTIALLY OVERLAND THE TWO, 

1479
00:52:20,796 --> 00:52:22,932
HOW MANY PIXELS WOULD BE OFF.

1480
00:52:22,932 --> 00:52:25,534
AND THAT NUMBER WAS LESS THAN 

1481
00:52:25,534 --> 00:52:27,570
ONE PIXEL WAS OFF.

1482
00:52:27,570 --> 00:52:29,405
SO IN THE NORMALIZED SCALE. 

1483
00:52:29,405 --> 00:52:30,306
>> OKAY, OKAY.

1484
00:52:30,306 --> 00:52:31,707
I'VE GOT THE IDEA, THANK YOU. 

1485
00:52:31,707 --> 00:52:32,475
>> YEAH, SURE. 

1486
00:52:32,475 --> 00:52:36,979
>> ACTUALLY, ONE MORE
QUESTION.

1487
00:52:36,979 --> 00:52:38,381
AND AT THE BEGINNING OF THE 

1488
00:52:38,381 --> 00:52:42,385
TALK, YOU LISTED A SET OF 

1489
00:52:42,385 --> 00:52:44,320
REASONS WHY ONE MAY NEED 

1490
00:52:44,320 --> 00:52:44,787
SYNTHETIC IMAGES. 

1491
00:52:44,787 --> 00:52:45,788
>> RIGHT. 

1492
00:52:45,788 --> 00:52:49,959
>> YOU MENTIONED
INCREASING 

1493
00:52:49,959 --> 00:52:51,861
NOISE, INCREASE IN THE VARIETY 

1494
00:52:51,861 --> 00:52:52,261
OF YOUR DATA. 

1495
00:52:52,261 --> 00:52:53,129
>> CORRECT. 

1496
00:52:53,129 --> 00:52:56,732
>> REDUCING THE DATA
IMBALANCE.

1497
00:52:56,732 --> 00:52:58,134
ARE THERE ANY CONDITIONS WHERE 

1498
00:52:58,134 --> 00:53:00,169
YOU WOULD LIKE TO GENERATE 

1499
00:53:00,169 --> 00:53:01,637
SYNTHETIC IMAGE BECAUSE YOU 

1500
00:53:01,637 --> 00:53:03,606
COULD NOT PRODUCE IT 

1501
00:53:03,606 --> 00:53:04,306
EXPERIMENTALLY FOR SOME 

1502
00:53:04,306 --> 00:53:04,774
TECHNICAL REASONS?

1503
00:53:04,774 --> 00:53:07,076
>> THAT IS A VERY VALID
POINT.

1504
00:53:07,076 --> 00:53:08,377
IT HASN'T FALLEN WITHIN MY 

1505
00:53:08,377 --> 00:53:11,013
PURVIEW BECAUSE IN MY CASE, IT'S

1506
00:53:11,013 --> 00:53:12,415
USUALLY DIAGNOSTIC, SO I'M 

1507
00:53:12,415 --> 00:53:14,150
TRYING TO DETECT DISEASE 

1508
00:53:14,150 --> 00:53:17,153
PATTERNS, BUT YES, IF YOU WANT 

1509
00:53:17,153 --> 00:53:20,990
TO JUST SYNTHESIZE SOMETHING, I 

1510
00:53:20,990 --> 00:53:22,091
MEAN, BASICALLY FORGET MEDICINE 

1511
00:53:22,091 --> 00:53:23,859
FOR A SECOND.

1512
00:53:23,859 --> 00:53:29,331
IF YOU GO TO GEMINI OR STABLE 

1513
00:53:29,331 --> 00:53:30,866
DIFFUSION OF THE DALLY TREE, IT 

1514
00:53:30,866 --> 00:53:32,468
IS DOING JUST THAT.

1515
00:53:32,468 --> 00:53:33,869
IT SYNTHESIZING IMAGES OUT OF 

1516
00:53:33,869 --> 00:53:34,103
CONTEXT.

1517
00:53:34,103 --> 00:53:35,237
THE ASSUMPTION YOU'RE MAKING IS 

1518
00:53:35,237 --> 00:53:37,907
THAT THE MODEL FROM WHICH YOU'RE

1519
00:53:37,907 --> 00:53:39,608
SYNTHESIZING HAS THE WORDS 

1520
00:53:39,608 --> 00:53:41,444
EMBEDDED IN THERE THAT BELONG TO

1521
00:53:41,444 --> 00:53:44,914
THE -- TO WHAT YOU WANT TO SEE 

1522
00:53:44,914 --> 00:53:45,948
OUT OF IT.

1523
00:53:45,948 --> 00:53:48,117
AND, TWO, THAT YOU HAVE A SCRIM 

1524
00:53:48,117 --> 00:53:48,984
NATOR OF SOME KIND, WHETHER 

1525
00:53:48,984 --> 00:53:51,720
HUMAN OR MACHINE, THAT WILL 

1526
00:53:51,720 --> 00:53:52,955
DETERMINE THAT WHAT HAS BEEN 

1527
00:53:52,955 --> 00:53:54,056
SYNTHESIZED BELONGS TO SOMETHING

1528
00:53:54,056 --> 00:53:57,259
YOU WOULD LIKE TO SEE AN IMAGE 

1529
00:53:57,259 --> 00:53:57,426
OF.

1530
00:53:57,426 --> 00:53:58,194
FROM A VISUALIZATION 

1531
00:53:58,194 --> 00:53:59,995
PERSPECTIVE, IT'S FANTASTIC, BUT

1532
00:53:59,995 --> 00:54:01,964
THAT FALLS MORE IN THE ART 

1533
00:54:01,964 --> 00:54:04,300
CATEGORY, BUT FROM A -- FROM THE

1534
00:54:04,300 --> 00:54:04,934
APPLICATION SPACE THAT I WAS 

1535
00:54:04,934 --> 00:54:06,502
TALKING ABOUT WHERE YOU ARE 

1536
00:54:06,502 --> 00:54:08,137
TRYING TO DETECT DISEASE AND 

1537
00:54:08,137 --> 00:54:09,638
CORRECT FOR THE IMBALANCE, IT 

1538
00:54:09,638 --> 00:54:11,140
DOESN'T FIT VERY WELL IN THAT 

1539
00:54:11,140 --> 00:54:11,507
IDEA.

1540
00:54:11,507 --> 00:54:14,310
BUT AS AN IDEA, THAT EXISTS.

1541
00:54:14,310 --> 00:54:15,177
THE CHALLENGE BECOMES ARE THE 

1542
00:54:15,177 --> 00:54:16,579
WORDS THAT YOU'RE GOING TO USE 

1543
00:54:16,579 --> 00:54:18,714
TO SYNTHESIZE SOMETHING FROM 

1544
00:54:18,714 --> 00:54:20,916
NOTHING, DO THEY BELONG TO THE 

1545
00:54:20,916 --> 00:54:25,621
TEXT VOCABULARY THAT THE 

1546
00:54:25,621 --> 00:54:28,357
DIFFUSION MODEL/LLM WAS EXPOSED 

1547
00:54:28,357 --> 00:54:31,427
TO WITH THE RIGHT KIND OF 

1548
00:54:31,427 --> 00:54:31,794
APPEARANCES.

1549
00:54:31,794 --> 00:54:33,362
YOU HAVE TO DESIGN SOME SIGNAL 

1550
00:54:33,362 --> 00:54:34,964
FROM SOMEWHERE TO SYNTHESIZE.

1551
00:54:34,964 --> 00:54:37,166
IT HAS TO BE EXPOSED, MAYBE NOT 

1552
00:54:37,166 --> 00:54:38,601
SPECIFICALLY TO A DATASET LIKE 

1553
00:54:38,601 --> 00:54:40,903
IN OUR CASE, BUT IN THE BROAD 

1554
00:54:40,903 --> 00:54:42,705
WORLD, IT WOULD JUST -- WE'VE 

1555
00:54:42,705 --> 00:54:44,173
SEEN SOME EXAMPLES WHICH WOULD 

1556
00:54:44,173 --> 00:54:45,407
RELATE TO THE WORDS THAT YOU'RE 

1557
00:54:45,407 --> 00:54:47,042
USING FOR CORE SYNTHESIZING MAP 

1558
00:54:47,042 --> 00:54:48,811
THE APPEARANCE THAT YOU EXPECT 

1559
00:54:48,811 --> 00:54:49,578
TO SEE.

1560
00:54:49,578 --> 00:54:51,046
NOW, IF IT'S JUST PRETTY GRAPHS 

1561
00:54:51,046 --> 00:54:52,615
AND THINGS LIKE THAT, I MEAN, 

1562
00:54:52,615 --> 00:54:54,116
THERE ARE SITES THAT WILL CREATE

1563
00:54:54,116 --> 00:54:56,752
CARTOONS BASED ON A PIECE OF 

1564
00:54:56,752 --> 00:54:56,952
TEXT.

1565
00:54:56,952 --> 00:54:59,188
I MEAN, IT'S A STYLE GAN 

1566
00:54:59,188 --> 00:55:01,290
UNDERNEATH SOMEWHERE WITH AN LDM

1567
00:55:01,290 --> 00:55:04,627
MODEL WHICH WOULD SAY I WANT A 

1568
00:55:04,627 --> 00:55:07,196
CARTOON IN THE STYLE OF DISNEY 

1569
00:55:07,196 --> 00:55:09,899
FOR EXAMPLE AND IT CREATES A 

1570
00:55:09,899 --> 00:55:10,833
CARTOON IN THE RECONCILE OF 

1571
00:55:10,833 --> 00:55:12,535
DISNEY CHARACTERS, BUT THE 

1572
00:55:12,535 --> 00:55:13,969
CONTENT OF THE CARTOON IS 

1573
00:55:13,969 --> 00:55:16,238
COMPLETELY DRIVEN BY YOU AND THE

1574
00:55:16,238 --> 00:55:18,040
FUNCTION OF THE WORDS IN THE 

1575
00:55:18,040 --> 00:55:18,407
CONTEXT.

1576
00:55:18,407 --> 00:55:19,408
I TRIED THIS YESTERDAY.

1577
00:55:19,408 --> 00:55:21,677
I FORGET WHAT SITE IT WAS, JUST 

1578
00:55:21,677 --> 00:55:23,779
TO SEE SOME EXAMPLES, AND IT 

1579
00:55:23,779 --> 00:55:25,114
FAILED MISERABLY.

1580
00:55:25,114 --> 00:55:27,349
THERE IS FOR WORD SCIENTIST, 

1581
00:55:27,349 --> 00:55:28,450
THERE IS NO WORD MEDICINE, THERE

1582
00:55:28,450 --> 00:55:31,987
IS NO WORD X-RAY, SO THE 

1583
00:55:31,987 --> 00:55:34,390
VOCABULARY IS PRETTY -- PRETTY 

1584
00:55:34,390 --> 00:55:36,258
CHILD CARTOON LIKE, BALLOONS AND

1585
00:55:36,258 --> 00:55:39,728
ALL THAT CAME THROUGH JUST FINE.

1586
00:55:39,728 --> 00:55:42,798
SO IT'S -- IT'S WHAT -- LIKE MY 

1587
00:55:42,798 --> 00:55:44,066
CLOSING SLIDE, WHAT IS THE 

1588
00:55:44,066 --> 00:55:45,668
CONTEXT IN WHICH YOU'RE TRYING 

1589
00:55:45,668 --> 00:55:46,969
TO SYNTHESIZE AND IT DOES MAKE 

1590
00:55:46,969 --> 00:55:47,870
SENSE IN THAT CONTEXT. 

1591
00:55:47,870 --> 00:55:49,705
>> OKAY, THANK YOU.

1592
00:55:49,705 --> 00:55:50,306
>> SURE.

1593
00:55:50,306 --> 00:55:52,474
>> OKAY, WE HAVE A
QUESTION FROM

1594
00:55:52,474 --> 00:55:53,709
THE VIDEO CAST.

1595
00:55:53,709 --> 00:55:55,811
THE QUESTION SAYS, ALTHOUGH I 

1596
00:55:55,811 --> 00:55:59,148
BELIEVE YOU ONLY USE THE 

1597
00:55:59,148 --> 00:56:01,517
GENERATOR FROM A GAN, 

1598
00:56:01,517 --> 00:56:02,351
EXAMINATIONING THE DISCRIMINATOR

1599
00:56:02,351 --> 00:56:04,053
MAY HELP YOU UNDERSTAND ITS 

1600
00:56:04,053 --> 00:56:04,353
LIMITATIONS.

1601
00:56:04,353 --> 00:56:05,354
DO YOU KNOW, FOR EXAMPLE, 

1602
00:56:05,354 --> 00:56:08,090
WHETHER THE DISCRIMINATOR 

1603
00:56:08,090 --> 00:56:09,024
RECORRECTS EXAMPLES THAT WERE 

1604
00:56:09,024 --> 00:56:10,993
NOT IN THE TRAINING -- REJECTS 

1605
00:56:10,993 --> 00:56:12,027
EXAMPLES THAT WERE NOT IN THE 

1606
00:56:12,027 --> 00:56:12,928
TRAINING SET?

1607
00:56:12,928 --> 00:56:14,363
>> THAT'S A GOOD QUESTIONS
ABOUT

1608
00:56:14,363 --> 00:56:15,231
WE HAVEN'T EXPLORED THAT.

1609
00:56:15,231 --> 00:56:17,566
I THINK WHAT THE QUESTION IS 

1610
00:56:17,566 --> 00:56:19,902
TRYING TO POINT AT IS WAS THE 

1611
00:56:19,902 --> 00:56:23,339
LOSS FUNCTION ADEQUATE IN -- WAS

1612
00:56:23,339 --> 00:56:26,375
IT TOO STRICT IN REJECTING GOOD 

1613
00:56:26,375 --> 00:56:29,845
EXAMPLES THAT DIDN'T FIT THE 

1614
00:56:29,845 --> 00:56:30,846
DISCRIMINATOR MODEL.

1615
00:56:30,846 --> 00:56:32,848
IT'S TIED TO THE REAL DATA IT 

1616
00:56:32,848 --> 00:56:36,452
HAS SEEN TO TRAIN IT, AND I'M 

1617
00:56:36,452 --> 00:56:38,087
NOT SO SURE IF -- WE HAVEN'T 

1618
00:56:38,087 --> 00:56:40,356
DONE, THIS OF COURSE, BUT IT'S 

1619
00:56:40,356 --> 00:56:42,324
AN INTERESTING POINTER TO SEE 

1620
00:56:42,324 --> 00:56:47,229
WHAT ARE THE REJECTS OR THE 

1621
00:56:47,229 --> 00:56:50,399
PATHWAY FROM THE DISCRIMINATOR 

1622
00:56:50,399 --> 00:56:53,902
IN TERMS OF WHAT WOULD BE A -- 

1623
00:56:53,902 --> 00:56:55,170
WHAT COULD IT HAVE BEEN, SEE 

1624
00:56:55,170 --> 00:56:56,572
THAT POOL AND SEE IF THERE WERE 

1625
00:56:56,572 --> 00:56:58,607
ANY CANDIDATES BECAUSE, AGAIN, 

1626
00:56:58,607 --> 00:57:01,910
RELYING ON MACHINE THRESHOLDS TO

1627
00:57:01,910 --> 00:57:03,846
DETERMINE THE LOSS THRESHOLD TO 

1628
00:57:03,846 --> 00:57:07,716
DETERMINE IF AN IMAGE PASSES THE

1629
00:57:07,716 --> 00:57:09,184
DISCRIMINATOR STAGE INTO THE 

1630
00:57:09,184 --> 00:57:10,419
OUTPUT, AND IF THAT THRESHOLD IS

1631
00:57:10,419 --> 00:57:12,254
TOO TIGHT, YOU'RE GOING TO GET 

1632
00:57:12,254 --> 00:57:14,156
THINGS THAT ARE GOOD FIT, BUT 

1633
00:57:14,156 --> 00:57:18,160
MAY BE COSTING US, THE VARIETY.

1634
00:57:18,160 --> 00:57:19,695
BUT THEN THE CHALLENGE BECOMES, 

1635
00:57:19,695 --> 00:57:21,030
WHICH IS A GOOD THRESHOLD IN THE

1636
00:57:21,030 --> 00:57:23,699
FIRST PLACE SOIRNTION A BIT OF A

1637
00:57:23,699 --> 00:57:25,768
CHICKEN AND EGG -- SO IT'S A BIT

1638
00:57:25,768 --> 00:57:27,236
OF A CHICKEN AND EGG, BUT FROM A

1639
00:57:27,236 --> 00:57:28,237
STUDY PERSPECTIVE, AN 

1640
00:57:28,237 --> 00:57:29,838
INTERESTING THING TO SEE.

1641
00:57:29,838 --> 00:57:32,908
HOW DOES ONE ARRIVE AT A 

1642
00:57:32,908 --> 00:57:36,211
MEANINGFUL LOSS THRESHOLD AT THE

1643
00:57:36,211 --> 00:57:37,680
DISCRIMINATOR SO YOU CAN INDUCE 

1644
00:57:37,680 --> 00:57:38,614
THE VARIETY YOU WANT WITHOUT 

1645
00:57:38,614 --> 00:57:40,182
COMPENSATING FOR QUALITY.

1646
00:57:40,182 --> 00:57:40,516
>> ALL RIGHT.

1647
00:57:40,516 --> 00:57:42,351
THANK YOU.

1648
00:57:42,351 --> 00:57:46,722
ANYMORE QUESTIONS HERE?

1649
00:57:46,722 --> 00:57:46,922
OKAY.

1650
00:57:46,922 --> 00:57:50,359
I THINK I JUST HAVE A FINAL 

1651
00:57:50,359 --> 00:57:50,626
QUESTION.

1652
00:57:50,626 --> 00:57:53,796
I BELIEVE IN ONE OF THE SLIDES, 

1653
00:57:53,796 --> 00:57:55,764
YOU WERE COMPARING GAN VERSUS 

1654
00:57:55,764 --> 00:57:57,833
JUST TRADITIONAL AUGMENTATION, 

1655
00:57:57,833 --> 00:57:59,301
AND CAN YOU ELABORATE A LITTLE 

1656
00:57:59,301 --> 00:58:02,237
BIT MORE ON THE METHOD THAT YOU 

1657
00:58:02,237 --> 00:58:03,872
USED FOR COMPARING BETWEEN THOSE

1658
00:58:03,872 --> 00:58:04,440
TWO?

1659
00:58:04,440 --> 00:58:05,674
I MEAN, HOW DO YOU KNOW WHICH 

1660
00:58:05,674 --> 00:58:07,976
ONE IS PERFORMING BETTER?

1661
00:58:07,976 --> 00:58:11,280
>> THE OUTPUT OF A
CLASSIFIER.

1662
00:58:11,280 --> 00:58:14,083
SO YOUR TRADITIONAL 

1663
00:58:14,083 --> 00:58:14,950
AUGMENTATION, YOU HAVE 

1664
00:58:14,950 --> 00:58:16,151
ESSENTIALLY A DATASET THAT IS 

1665
00:58:16,151 --> 00:58:19,221
TRADITIONAL AUGMENTED AND THAT 

1666
00:58:19,221 --> 00:58:21,957
IS THE BLOCK HERE, SO YOU'RE 

1667
00:58:21,957 --> 00:58:24,059
TRAINING ONE AND YOU'RE 

1668
00:58:24,059 --> 00:58:24,960
COMPARING AGAINST -- YOU'RE 

1669
00:58:24,960 --> 00:58:27,096
TRAINING AGAINST A TRADITIONAL 

1670
00:58:27,096 --> 00:58:30,399
AUGMENTATION IMAGE DATASET, 

1671
00:58:30,399 --> 00:58:33,936
WHICH IS -- A GAN-BASED DATASET 

1672
00:58:33,936 --> 00:58:35,537
AND SEPARATE DATASET AND THE 

1673
00:58:35,537 --> 00:58:37,373
ORIGINAL UNTOUCHED DATASET, SO 

1674
00:58:37,373 --> 00:58:38,374
YOU HAVE THREE THINGS COMPARED 

1675
00:58:38,374 --> 00:58:39,808
AGAINST EACH OTHER WHICH BECOME 

1676
00:58:39,808 --> 00:58:40,776
THE COMPARISON ON THIS TABLE 

1677
00:58:40,776 --> 00:58:42,611
WHERE THE FIRST ROW IS NO 

1678
00:58:42,611 --> 00:58:44,980
AUGMENTATION OF ANY KIND AND 

1679
00:58:44,980 --> 00:58:47,149
THAT'S THE BASELINE PERFORMANCE.

1680
00:58:47,149 --> 00:58:47,916
THIS IS WHAT SOMEBODY MIGHT JUST

1681
00:58:47,916 --> 00:58:51,620
DO OFF THE BAT, TRADITIONAL 

1682
00:58:51,620 --> 00:58:53,255
AUGMENTATION AND ALL THE FINE 

1683
00:58:53,255 --> 00:58:53,455
STUFF.

1684
00:58:53,455 --> 00:58:56,558
AND YOU GET NUMBERS THAT ARE 

1685
00:58:56,558 --> 00:58:58,394
OBVIOUSLY BETTER -- IN THE GA, 

1686
00:58:58,394 --> 00:58:59,728
WE DID NOT DO ANY AUGMENTATION 

1687
00:58:59,728 --> 00:59:01,697
THAT IS TRADITIONALLY DONE.

1688
00:59:01,697 --> 00:59:04,166
WE JUST INCREASED THE TEST SIZE.

1689
00:59:04,166 --> 00:59:06,635
SO GA IN THIS ROW IS A LOT LIKE 

1690
00:59:06,635 --> 00:59:08,370
NO AUGMENTATION, EXCEPT THE 

1691
00:59:08,370 --> 00:59:09,905
NUMBER OF IMAGES HAVE GONE UP 

1692
00:59:09,905 --> 00:59:10,406
AND BAM OUT.

1693
00:59:10,406 --> 00:59:12,174
SO YOU BALANCE -- IN THIS CASE, 

1694
00:59:12,174 --> 00:59:14,343
IT'S A HIGHLY I AM BALANCED SET,

1695
00:59:14,343 --> 00:59:15,177
THE FIRST ROW.

1696
00:59:15,177 --> 00:59:17,146
IN THE THIRD ROW, IT'S A 

1697
00:59:17,146 --> 00:59:19,648
BALANCED SET, BUT WITH MORE 

1698
00:59:19,648 --> 00:59:21,283
SAMPLES THAT WERE SYNTHESIZED 

1699
00:59:21,283 --> 00:59:23,919
AND THE MIDDLE ONE IS IN PROCESS

1700
00:59:23,919 --> 00:59:26,722
AUGMENTATION THAT IS DONE TO 

1701
00:59:26,722 --> 00:59:27,556
CORRECT FOR THE IMBALANCE 

1702
00:59:27,556 --> 00:59:29,324
THAT -- THE TRADITIONAL 

1703
00:59:29,324 --> 00:59:30,759
TECHNIQUE AS ANYBODY ELSE MIGHT 

1704
00:59:30,759 --> 00:59:30,926
DO.

1705
00:59:30,926 --> 00:59:33,896
SO WE WANTED TO SEE THE EFFECT 

1706
00:59:33,896 --> 00:59:38,400
OF ADDING JUST MORE IMAGES FOR 

1707
00:59:38,400 --> 00:59:39,101
THE AUGMENTATION PART.

1708
00:59:39,101 --> 00:59:41,503
OUR AUGMENTATION STEP IS 

1709
00:59:41,503 --> 00:59:42,371
SYNTHESIS.

1710
00:59:42,371 --> 00:59:43,739
NOW, ONE CAN GO FURTHER.

1711
00:59:43,739 --> 00:59:45,140
TAKE THE GA IMAGES AND DO 

1712
00:59:45,140 --> 00:59:46,275
TRADITIONAL AUGMENTATION ON ALL 

1713
00:59:46,275 --> 00:59:47,576
OF THEM.

1714
00:59:47,576 --> 00:59:49,178
NOTHING WRONG IN DOING THAT, AND

1715
00:59:49,178 --> 00:59:50,712
MAYBE THE NUMBER WOULD BE 

1716
00:59:50,712 --> 00:59:51,413
SLIGHTLY BETTER.

1717
00:59:51,413 --> 00:59:54,349
WHO KNOWS, BUT THE POINT WAS TO 

1718
00:59:54,349 --> 00:59:57,786
ACTUALLY SEE WHAT DID GENERATIVE

1719
00:59:57,786 --> 01:00:00,722
AI BRING ANY VALUE, PLAIN 

1720
01:00:00,722 --> 01:00:02,090
VANILLA, AND IT JUST -- IT SAYS 

1721
01:00:02,090 --> 01:00:04,660
IT DOESN'T BY LOOKING AT THE 

1722
01:00:04,660 --> 01:00:05,127
DATASET.

1723
01:00:05,127 --> 01:00:06,895
BUT IF YOU GOT THE TWO EXAMPLES 

1724
01:00:06,895 --> 01:00:07,996
AND YOU HAVE SEEN A SIMILAR 

1725
01:00:07,996 --> 01:00:10,232
TREND LINE WHERE WE ARE NOT 

1726
01:00:10,232 --> 01:00:11,934
SEEING A BUMP THAT WE HOPE TO 

1727
01:00:11,934 --> 01:00:16,238
SEE BY CORRECTING IMBALANCES..B 

1728
01:00:16,238 --> 01:00:16,472
FACTOR.

1729
01:00:16,472 --> 01:00:17,272
IT'S SOMETHING MORE THAN 

1730
01:00:17,272 --> 01:00:19,308
IMBALANCE IS THE FACTOR.

1731
01:00:19,308 --> 01:00:21,710
SO IMBALANCE -- IF YOU LOOK AT 

1732
01:00:21,710 --> 01:00:23,412
JUST AUGMENTATION AND NO 

1733
01:00:23,412 --> 01:00:25,814
AUGMENTATION, IMBALANCE PLAYS A 

1734
01:00:25,814 --> 01:00:26,215
ROLE, NO DOUBT.

1735
01:00:26,215 --> 01:00:29,384
NOT A BIG ONE IN THIS, NOT SO 

1736
01:00:29,384 --> 01:00:34,089
AMPLIFIED IN CASE, BUT IT DOES 

1737
01:00:34,089 --> 01:00:38,126
PLAY A .. GENERATIVE AI IS NOT Y

1738
01:00:38,126 --> 01:00:38,360
PATHWAY.

1739
01:00:38,360 --> 01:00:40,596
IT MAY BE ONE OF THE PATHWAYS 

1740
01:00:40,596 --> 01:00:41,797
COMBINED WITH SOMETHING ELSE AND

1741
01:00:41,797 --> 01:00:43,632
THAT'S THE WHOLE POINT OF THIS 

1742
01:00:43,632 --> 01:00:45,234
EXPERIMENT WAT TO INDUCE THAT 

1743
01:00:45,234 --> 01:00:45,801
PART.

1744
01:00:45,801 --> 01:00:46,502
JUST BECAUSE YOU HAVE A 

1745
01:00:46,502 --> 01:00:48,804
MECHANISM TO SYNTHESIZE NEW DATA

1746
01:00:48,804 --> 01:00:50,038
DOESN'T MEAN -- OF COURSE, 

1747
01:00:50,038 --> 01:00:50,739
THERE'S A WHOLE SEPARATE FIELD 

1748
01:00:50,739 --> 01:00:52,875
THAT IS DOING SYNTHESIZING DATA 

1749
01:00:52,875 --> 01:00:56,378
ON TABULAR DATA WHERE YOU'RE 

1750
01:00:56,378 --> 01:01:00,115
MISSING FIELDS, MISSING VALUES, 

1751
01:01:00,115 --> 01:01:03,785
AND YOU GET INTO SIMILAR  

1752
01:01:03,785 --> 01:01:06,455
PROBLEMS, BU DO THE PERSON'S 

1753
01:01:06,455 --> 01:01:08,257
SIZE AND WEIGHT MAKE A 

1754
01:01:08,257 --> 01:01:09,224
DIFFERENCE, DO THE BLOOD VALUES 

1755
01:01:09,224 --> 01:01:10,859
MAKE SENSE OR NOT, ETCETERA.

1756
01:01:10,859 --> 01:01:14,096
SO THE POINT IS TO GET YOUR FEET

1757
01:01:14,096 --> 01:01:16,431
WET AND RECOGNIZE THAT IT'S A 

1758
01:01:16,431 --> 01:01:19,434
VIABLE TECHNOLOGY, MEANINGFUL 

1759
01:01:19,434 --> 01:01:21,670
WORK USING -- GIVEN THE 

1760
01:01:21,670 --> 01:01:22,271
REALITIES THAT WE FACE WITH 

1761
01:01:22,271 --> 01:01:24,773
CHALLENGES.

1762
01:01:24,773 --> 01:01:25,274
>> GREAT.

1763
01:01:25,274 --> 01:01:26,542
THANK YOU FOR THE CLARIFICATION.

1764
01:01:26,542 --> 01:01:28,143
AND THANK YOU FOR A GREAT TALK.

1765
01:01:28,143 --> 01:01:38,353
[ APPLAUSE ]