1
00:00:04,960 --> 00:00:19,999
[Music]

2
00:00:20,439 --> 00:00:25,359
and with us we have Jay Rosen bomb who's

3
00:00:23,080 --> 00:00:27,599
a Melbourne based AI artist and

4
00:00:25,359 --> 00:00:29,679
researcher working in 3D modeling

5
00:00:27,599 --> 00:00:32,960
artificial intelligence and extended

6
00:00:29,679 --> 00:00:35,719
real Technologies they have a PhD from

7
00:00:32,960 --> 00:00:38,239
rmit University uh studying AI

8
00:00:35,719 --> 00:00:41,920
perception of genders and the nature of

9
00:00:38,239 --> 00:00:44,719
AI generated art and human gen uh human

10
00:00:41,920 --> 00:00:46,879
hands behind the process that engender

11
00:00:44,719 --> 00:00:49,520
the bias uh and today they are going to

12
00:00:46,879 --> 00:00:51,840
give us a talk on AI perceptions of

13
00:00:49,520 --> 00:00:52,770
gender I'll hand it over to you thank

14
00:00:51,840 --> 00:00:57,879
you so

15
00:00:52,770 --> 00:01:01,039
[Applause]

16
00:00:57,879 --> 00:01:05,479
much yep just wait for the slides to pop

17
00:01:01,039 --> 00:01:08,080
but yeah so hi hi everyone um I'm so I'm

18
00:01:05,479 --> 00:01:10,240
so excited to be here today I love Pon

19
00:01:08,080 --> 00:01:12,520
uh my name is Jay Rosen and I'm an AI

20
00:01:10,240 --> 00:01:15,320
artist and researcher and I te and I'm a

21
00:01:12,520 --> 00:01:16,520
lecturer in critical AI at rmit school

22
00:01:15,320 --> 00:01:19,720
of

23
00:01:16,520 --> 00:01:23,360
design the term AI artist has become a

24
00:01:19,720 --> 00:01:25,280
really loaded term lately um when I

25
00:01:23,360 --> 00:01:29,479
first started working with machine

26
00:01:25,280 --> 00:01:31,799
learning art nearly a decade ago it was

27
00:01:29,479 --> 00:01:34,880
easier to to say AI artist The Machine

28
00:01:31,799 --> 00:01:36,840
learning artist um but nobody knew what

29
00:01:34,880 --> 00:01:40,640
machine learning meant an AI artist was

30
00:01:36,840 --> 00:01:43,799
sort of a Sci-Fi fancy fun marketing

31
00:01:40,640 --> 00:01:46,840
term and now of course it's vokes ideas

32
00:01:43,799 --> 00:01:48,600
of shortcuts and plagiarism and creative

33
00:01:46,840 --> 00:01:50,960
theft and I don't really like being

34
00:01:48,600 --> 00:01:53,040
associated with that anymore but I do

35
00:01:50,960 --> 00:01:54,880
still use the term after a lot of

36
00:01:53,040 --> 00:01:56,360
internal wrangling I do still use the

37
00:01:54,880 --> 00:01:58,439
term AI

38
00:01:56,360 --> 00:02:01,840
artist sometimes with a little bit of

39
00:01:58,439 --> 00:02:04,840
irony and because my practice deeply

40
00:02:01,840 --> 00:02:09,319
engages with AI systems not just as

41
00:02:04,840 --> 00:02:11,599
tools but as objects of critique and

42
00:02:09,319 --> 00:02:13,560
study today I'm going to talk about my

43
00:02:11,599 --> 00:02:16,360
research uh and the art I created during

44
00:02:13,560 --> 00:02:19,680
my PhD which explored AI perceptions of

45
00:02:16,360 --> 00:02:22,280
gender my focus was on how AI interprets

46
00:02:19,680 --> 00:02:24,080
and represents gender challenging the

47
00:02:22,280 --> 00:02:26,080
binary Frameworks that dominate machine

48
00:02:24,080 --> 00:02:27,840
learning and expanding them Beyond

49
00:02:26,080 --> 00:02:31,400
biological

50
00:02:27,840 --> 00:02:34,120
essentialism gender is fast it's deeply

51
00:02:31,400 --> 00:02:37,519
personal it's so much more than just

52
00:02:34,120 --> 00:02:39,959
Anatomy or a checkbox on a form through

53
00:02:37,519 --> 00:02:42,400
my work I've asked questions like why do

54
00:02:39,959 --> 00:02:44,319
AI systems perceive gender and how do

55
00:02:42,400 --> 00:02:47,760
they qualify gender in the images that

56
00:02:44,319 --> 00:02:50,840
they generate if we create a biased data

57
00:02:47,760 --> 00:02:53,360
set and slowly debias it by introducing

58
00:02:50,840 --> 00:02:55,599
new data how will the generated images

59
00:02:53,360 --> 00:02:58,319
be affected and what are the

60
00:02:55,599 --> 00:03:01,040
implications for visual representation

61
00:02:58,319 --> 00:03:02,959
present in generative images of people a

62
00:03:01,040 --> 00:03:06,879
lot of these questions are still ongoing

63
00:03:02,959 --> 00:03:08,519
today like the PHD is finished but I'm

64
00:03:06,879 --> 00:03:11,599
still finding myself coming back to

65
00:03:08,519 --> 00:03:14,040
these questions over and over

66
00:03:11,599 --> 00:03:16,080
again I trained a bias neural network

67
00:03:14,040 --> 00:03:17,720
and I sought to see if I could add data

68
00:03:16,080 --> 00:03:20,120
to change its biases and what it looks

69
00:03:17,720 --> 00:03:22,080
like if I did I examined existing

70
00:03:20,120 --> 00:03:24,959
systems and looked at how they explore

71
00:03:22,080 --> 00:03:27,640
gender and what they see when presented

72
00:03:24,959 --> 00:03:29,319
with gender beyond the binary and I

73
00:03:27,640 --> 00:03:30,360
explored a new way of looking at gender

74
00:03:29,319 --> 00:03:33,480
classific

75
00:03:30,360 --> 00:03:35,159
completely by assigning people a custom

76
00:03:33,480 --> 00:03:38,120
mixed

77
00:03:35,159 --> 00:03:40,840
color if you were at Pon last year you

78
00:03:38,120 --> 00:03:43,519
might have caught my talk on that one um

79
00:03:40,840 --> 00:03:45,480
my first project set in stone began with

80
00:03:43,519 --> 00:03:48,280
exploring whether we can change these

81
00:03:45,480 --> 00:03:50,360
biases in a biased system uh people

82
00:03:48,280 --> 00:03:52,840
often talk about the way to fix a biased

83
00:03:50,360 --> 00:03:55,560
system is to add more data so I went

84
00:03:52,840 --> 00:03:58,760
okay let's start from that point let's

85
00:03:55,560 --> 00:04:00,879
add more data let's change the narrative

86
00:03:58,760 --> 00:04:03,599
so I worked with generative adversarial

87
00:04:00,879 --> 00:04:05,480
networks Gans um using techniques like

88
00:04:03,599 --> 00:04:07,120
transfer learning where you train

89
00:04:05,480 --> 00:04:09,400
retrain a neural network after it's

90
00:04:07,120 --> 00:04:11,760
already been trained and what I liked to

91
00:04:09,400 --> 00:04:14,799
call disruption training where I would

92
00:04:11,760 --> 00:04:17,199
just add in new data throughout the

93
00:04:14,799 --> 00:04:20,560
training process while while it was

94
00:04:17,199 --> 00:04:23,120
still training um and that approach

95
00:04:20,560 --> 00:04:25,639
really allowed me to see whether a gan

96
00:04:23,120 --> 00:04:29,520
could unlearn its original biases and

97
00:04:25,639 --> 00:04:29,520
adapt to new representation

98
00:04:32,960 --> 00:04:36,600
sorry masks make you

99
00:04:34,480 --> 00:04:39,080
[Laughter]

100
00:04:36,600 --> 00:04:41,520
dry the light the title comes from a

101
00:04:39,080 --> 00:04:44,440
metaphor that I kept coming back to the

102
00:04:41,520 --> 00:04:47,120
way AI systems encode bias as if carving

103
00:04:44,440 --> 00:04:49,240
it in stone so I started with DC Gan and

104
00:04:47,120 --> 00:04:51,039
pran models which generate images by

105
00:04:49,240 --> 00:04:53,320
learning patterns from a training data

106
00:04:51,039 --> 00:04:56,320
set my goal was to see if I could

107
00:04:53,320 --> 00:04:59,960
reshape that stone by injecting data

108
00:04:56,320 --> 00:05:02,080
that defied stereotypes I created 3D r

109
00:04:59,960 --> 00:05:04,800
faces for my data set representing

110
00:05:02,080 --> 00:05:07,919
masculine feminine and non-binary traits

111
00:05:04,800 --> 00:05:10,680
with classical marble Aesthetics leaning

112
00:05:07,919 --> 00:05:12,520
into that set in stone

113
00:05:10,680 --> 00:05:15,000
metaphor I tried to make them as

114
00:05:12,520 --> 00:05:18,800
racially diverse as possible I created

115
00:05:15,000 --> 00:05:21,400
an animation with 3D rotation and used

116
00:05:18,800 --> 00:05:24,800
all the Stills as the um training data

117
00:05:21,400 --> 00:05:27,120
set um in this 3D rendered

118
00:05:24,800 --> 00:05:29,240
environment so for this project I first

119
00:05:27,120 --> 00:05:31,039
trained a series of DC gains or deep

120
00:05:29,240 --> 00:05:33,479
convol IAL generative adversarial

121
00:05:31,039 --> 00:05:36,560
networks with my data set of masculine

122
00:05:33,479 --> 00:05:39,919
faces over time when once it was well

123
00:05:36,560 --> 00:05:42,319
and truly biased I introduced feminine

124
00:05:39,919 --> 00:05:45,360
faces and then faces representing more

125
00:05:42,319 --> 00:05:48,160
non-binary Aesthetics images emphasizing

126
00:05:45,360 --> 00:05:50,680
external markers of gender expression

127
00:05:48,160 --> 00:05:52,400
like brightly colored hair and makeup

128
00:05:50,680 --> 00:05:54,520
these elements were chosen to highlight

129
00:05:52,400 --> 00:05:58,440
how gender minorities often communicate

130
00:05:54,520 --> 00:06:01,680
their uh their gender and highlight um

131
00:05:58,440 --> 00:06:04,319
how how we see our identity visually

132
00:06:01,680 --> 00:06:07,720
because non-binary people don't always

133
00:06:04,319 --> 00:06:11,479
you know look non-binary so how do you

134
00:06:07,720 --> 00:06:13,479
tell that to a computer how do you so

135
00:06:11,479 --> 00:06:14,919
it's a little on the nose the colored

136
00:06:13,479 --> 00:06:17,319
hair and makeup it's a bit of a

137
00:06:14,919 --> 00:06:19,919
stereotype but sometimes you need to be

138
00:06:17,319 --> 00:06:22,800
a little literal um and it worked really

139
00:06:19,919 --> 00:06:25,720
well for that breaking of the stone

140
00:06:22,800 --> 00:06:28,599
metaphor as new data entered the system

141
00:06:25,720 --> 00:06:30,960
the generative outputs shifted initially

142
00:06:28,599 --> 00:06:33,000
the gans's output reflected the rigid

143
00:06:30,960 --> 00:06:35,599
binary ideas of gender but as the

144
00:06:33,000 --> 00:06:38,919
training progressed these boundaries

145
00:06:35,599 --> 00:06:41,479
softened vibrant colors emerge and take

146
00:06:38,919 --> 00:06:43,120
and The Works take on a more fluid and

147
00:06:41,479 --> 00:06:45,599
joyful take on

148
00:06:43,120 --> 00:06:48,039
gender however even as the gain

149
00:06:45,599 --> 00:06:49,560
integrated the new data I noticed a

150
00:06:48,039 --> 00:06:52,120
tendency to revert to its original

151
00:06:49,560 --> 00:06:54,240
masculine bias over time increasing the

152
00:06:52,120 --> 00:06:57,520
volume of non-binary INF feminine coded

153
00:06:54,240 --> 00:07:00,039
data helped to counteract that reversion

154
00:06:57,520 --> 00:07:02,479
but the model never truly forgot its

155
00:07:00,039 --> 00:07:02,479
initial

156
00:07:04,240 --> 00:07:11,639
biases switching to a PR a pran or

157
00:07:08,400 --> 00:07:14,440
Progressive growing Gan yielded smoother

158
00:07:11,639 --> 00:07:16,960
transitions the technique used to debias

159
00:07:14,440 --> 00:07:18,960
was the key observation here as I had to

160
00:07:16,960 --> 00:07:21,840
create directories filled with the same

161
00:07:18,960 --> 00:07:24,800
masculine data set uh initially it's a

162
00:07:21,840 --> 00:07:27,759
labeled uh Gan so I needed to create the

163
00:07:24,800 --> 00:07:30,639
label data sets filled all with the same

164
00:07:27,759 --> 00:07:34,120
data and then changed the them up uh

165
00:07:30,639 --> 00:07:37,000
over time and then I so I replaced the

166
00:07:34,120 --> 00:07:38,800
data gradually with more diverse images

167
00:07:37,000 --> 00:07:41,080
this was a fitting metaphor for how some

168
00:07:38,800 --> 00:07:44,360
genders need to make room for the

169
00:07:41,080 --> 00:07:46,240
representation of others by gradually

170
00:07:44,360 --> 00:07:49,159
replacing older data sets with newer

171
00:07:46,240 --> 00:07:51,759
images I observed more seamless

172
00:07:49,159 --> 00:07:54,319
debiasing the pran created fewer

173
00:07:51,759 --> 00:07:56,759
artifacts and seemed better uh

174
00:07:54,319 --> 00:07:59,199
integrating diverse representations

175
00:07:56,759 --> 00:08:01,080
without deting to its original biases

176
00:07:59,199 --> 00:08:04,560
the resulting visuals were presented as

177
00:08:01,080 --> 00:08:06,440
mosaics wallpapers and looping videos

178
00:08:04,560 --> 00:08:10,400
from afar they appear as cohesive

179
00:08:06,440 --> 00:08:13,360
portraits but up close viewers could see

180
00:08:10,400 --> 00:08:15,560
that the hundreds of and thousands

181
00:08:13,360 --> 00:08:18,159
sometimes of smaller images that made up

182
00:08:15,560 --> 00:08:20,120
each face these layers reflected the

183
00:08:18,159 --> 00:08:21,599
idea that gender is shaped by

184
00:08:20,120 --> 00:08:24,319
accumulated experiences and

185
00:08:21,599 --> 00:08:26,520
self-expression over time through set in

186
00:08:24,319 --> 00:08:29,560
stone I learned that while debiasing

187
00:08:26,520 --> 00:08:31,759
again is possible it's not always

188
00:08:29,560 --> 00:08:35,399
straightforward the architecture of the

189
00:08:31,759 --> 00:08:37,440
neural network has to play a part in how

190
00:08:35,399 --> 00:08:41,120
successfully it can adapt to the new

191
00:08:37,440 --> 00:08:43,880
data and much like people AI systems can

192
00:08:41,120 --> 00:08:47,200
carry their history with them even as

193
00:08:43,880 --> 00:08:49,200
they evolve the large scale mosaics show

194
00:08:47,200 --> 00:08:51,480
not only the AI Aesthetics and the

195
00:08:49,200 --> 00:08:53,760
debiasing of the neural network but

196
00:08:51,480 --> 00:08:55,959
created a dialogue about how we are the

197
00:08:53,760 --> 00:08:57,800
sum of our experiences not just what you

198
00:08:55,959 --> 00:09:01,079
see on the

199
00:08:57,800 --> 00:09:02,920
surface in the loop video this one the

200
00:09:01,079 --> 00:09:04,959
face continually zooms into to one

201
00:09:02,920 --> 00:09:08,240
square of many in the Mosaic which then

202
00:09:04,959 --> 00:09:10,640
shifts and changes developing into a new

203
00:09:08,240 --> 00:09:12,959
face which is zoomed into again each

204
00:09:10,640 --> 00:09:15,640
face exists because of all the faces

205
00:09:12,959 --> 00:09:18,680
that surround it helping it

206
00:09:15,640 --> 00:09:21,320
become I show here not only the shift

207
00:09:18,680 --> 00:09:22,959
caused by debiasing and AI techniques

208
00:09:21,320 --> 00:09:25,800
but also the experience of being

209
00:09:22,959 --> 00:09:29,839
transgender the exploration of gender as

210
00:09:25,800 --> 00:09:31,760
observance as expression as experience

211
00:09:29,839 --> 00:09:34,200
this concept of gender as experience is

212
00:09:31,760 --> 00:09:36,880
one I returned to in my gender tapestry

213
00:09:34,200 --> 00:09:39,560
work discussing gender as a highly

214
00:09:36,880 --> 00:09:42,320
personal lived experience that is unique

215
00:09:39,560 --> 00:09:42,320
to every

216
00:09:46,079 --> 00:09:51,680
individual the second project in my PhD

217
00:09:49,160 --> 00:09:54,360
was an analysis of existing systems in a

218
00:09:51,680 --> 00:09:56,480
complex Loop of AI architectures if

219
00:09:54,360 --> 00:09:58,519
you've ever played a game of telephone

220
00:09:56,480 --> 00:10:01,680
you know how a message changes when it's

221
00:09:58,519 --> 00:10:02,760
passed along the chain or telestrations

222
00:10:01,680 --> 00:10:05,279
the board

223
00:10:02,760 --> 00:10:08,320
game I wanted to see what would happen

224
00:10:05,279 --> 00:10:11,240
if I turned that idea into an AI

225
00:10:08,320 --> 00:10:14,720
experiment what happens when I link a

226
00:10:11,240 --> 00:10:16,600
bunch of different AI systems together

227
00:10:14,720 --> 00:10:19,680
and observe their output so

228
00:10:16,600 --> 00:10:22,120
Frankenstein's telephone explored how AI

229
00:10:19,680 --> 00:10:24,839
systems gender images revealing the

230
00:10:22,120 --> 00:10:27,279
biases embedded in their data sets this

231
00:10:24,839 --> 00:10:29,959
project works like a game of telephone

232
00:10:27,279 --> 00:10:32,560
using a chain of AI systems that pass an

233
00:10:29,959 --> 00:10:33,399
image and its interpretation from one to

234
00:10:32,560 --> 00:10:36,920
the

235
00:10:33,399 --> 00:10:39,279
next the process started with attn Gan

236
00:10:36,920 --> 00:10:41,360
the very first text to image generator

237
00:10:39,279 --> 00:10:44,200
they created an image based on a user's

238
00:10:41,360 --> 00:10:47,600
input prompt that image was segmented by

239
00:10:44,200 --> 00:10:50,839
deeplab breaking it into labeled regions

240
00:10:47,600 --> 00:10:53,920
like Sky person tree and the segmented

241
00:10:50,839 --> 00:10:55,720
image was then reimagined by Spade Coco

242
00:10:53,920 --> 00:10:58,399
which reconstructed it as a new image

243
00:10:55,720 --> 00:11:00,720
and finally IM to text generated a

244
00:10:58,399 --> 00:11:03,120
caption for the fin image so that's

245
00:11:00,720 --> 00:11:04,800
probably very unreadable but that's okay

246
00:11:03,120 --> 00:11:06,480
cuz we've got a readable version coming

247
00:11:04,800 --> 00:11:09,440
up

248
00:11:06,480 --> 00:11:11,120
oh this chain highlighted how each

249
00:11:09,440 --> 00:11:14,680
system introduced new layers of

250
00:11:11,120 --> 00:11:17,040
interpretation Distortion and bias so

251
00:11:14,680 --> 00:11:19,920
this this what the prompt a woman with a

252
00:11:17,040 --> 00:11:23,639
dog on a leash it's now segmented the

253
00:11:19,920 --> 00:11:26,000
dog as partly um partly a

254
00:11:23,639 --> 00:11:28,200
person and um it's been a bit

255
00:11:26,000 --> 00:11:31,320
interesting which has then turned into

256
00:11:28,200 --> 00:11:35,160
this version this further version which

257
00:11:31,320 --> 00:11:38,240
um has a a these are very chaotic images

258
00:11:35,160 --> 00:11:40,160
these are are very early um generative

259
00:11:38,240 --> 00:11:43,920
AI systems they're not what we're used

260
00:11:40,160 --> 00:11:45,519
to now but then image to text um

261
00:11:43,920 --> 00:11:47,959
captioned it as a man writing a

262
00:11:45,519 --> 00:11:49,600
skateboard down a

263
00:11:47,959 --> 00:11:52,000
street

264
00:11:49,600 --> 00:11:53,560
so this chain highlighted how each

265
00:11:52,000 --> 00:11:55,920
system introduces new layers of

266
00:11:53,560 --> 00:11:58,560
interpretation Distortion and bias for

267
00:11:55,920 --> 00:12:01,000
example deep lab segmentation labeled

268
00:11:58,560 --> 00:12:04,120
all human simply as person removing

269
00:12:01,000 --> 00:12:05,959
gender entirely yet by the final stage

270
00:12:04,120 --> 00:12:08,800
gendered language had appeared in the

271
00:12:05,959 --> 00:12:11,440
captions again binary words like man or

272
00:12:08,800 --> 00:12:13,680
woman would emerge based on contextual

273
00:12:11,440 --> 00:12:16,199
cluthes like clothing or background

274
00:12:13,680 --> 00:12:18,160
elements I Trac this bias back to the

275
00:12:16,199 --> 00:12:20,560
human generated captions in the training

276
00:12:18,160 --> 00:12:22,360
data sets these captions frequently

277
00:12:20,560 --> 00:12:24,680
gendered people even when their faces

278
00:12:22,360 --> 00:12:26,959
weren't visible or when the person in

279
00:12:24,680 --> 00:12:29,160
question appeared abstracted or

280
00:12:26,959 --> 00:12:31,519
incomplete this pointed to the

281
00:12:29,160 --> 00:12:33,959
assumptions of the data set's human

282
00:12:31,519 --> 00:12:36,000
annotators many of whom worked under

283
00:12:33,959 --> 00:12:39,040
vague instructions through Mass labor

284
00:12:36,000 --> 00:12:40,920
platforms like Amazon Mechanical turque

285
00:12:39,040 --> 00:12:42,760
there's little to no training in

286
00:12:40,920 --> 00:12:45,639
situations like these just a short

287
00:12:42,760 --> 00:12:47,600
series of instructions it strikes me

288
00:12:45,639 --> 00:12:50,399
that a specialist could be hired to

289
00:12:47,600 --> 00:12:51,360
specify clearer instructions to avoid

290
00:12:50,399 --> 00:12:54,120
gendered

291
00:12:51,360 --> 00:12:56,160
outcomes like I observed in the um in

292
00:12:54,120 --> 00:12:58,440
the data set and especially the

293
00:12:56,160 --> 00:13:01,920
occasional sexist or racist outcomes I

294
00:12:58,440 --> 00:13:05,600
observed in the data set the um the data

295
00:13:01,920 --> 00:13:08,440
set often would uh have like a value

296
00:13:05,600 --> 00:13:10,240
judgments a very beautiful woman doing

297
00:13:08,440 --> 00:13:14,079
this and then if it was a man it was a

298
00:13:10,240 --> 00:13:16,199
man you know um whereas gender actually

299
00:13:14,079 --> 00:13:18,480
if one was just a hand and they went oh

300
00:13:16,199 --> 00:13:20,519
that's a woman's hand it's like it's a

301
00:13:18,480 --> 00:13:22,839
black and white hand it's barely in

302
00:13:20,519 --> 00:13:25,760
Focus who knows whose hand that is come

303
00:13:22,839 --> 00:13:29,360
on we don't need to gender these things

304
00:13:25,760 --> 00:13:32,160
but that gendering that bias that that

305
00:13:29,360 --> 00:13:34,120
came through those assumptions all came

306
00:13:32,160 --> 00:13:36,160
through in the

307
00:13:34,120 --> 00:13:38,440
work the visual outputs of

308
00:13:36,160 --> 00:13:41,720
Frankenstein's telephone are surreal and

309
00:13:38,440 --> 00:13:43,959
unsettling pale fleshy feature figures

310
00:13:41,720 --> 00:13:46,639
with exaggerated alien features these

311
00:13:43,959 --> 00:13:48,920
dreamlike blobs reflect the data set's

312
00:13:46,639 --> 00:13:50,759
lack of grounding in real human anatomy

313
00:13:48,920 --> 00:13:52,759
and raise questions about how machines

314
00:13:50,759 --> 00:13:55,800
create images and perceive images around

315
00:13:52,759 --> 00:13:57,920
gender for me the work became a metaphor

316
00:13:55,800 --> 00:13:59,360
for how transgender and non-gender and

317
00:13:57,920 --> 00:14:00,160
gender non-conforming people people are

318
00:13:59,360 --> 00:14:03,040
often

319
00:14:00,160 --> 00:14:04,440
perceived misunderstood distorted by

320
00:14:03,040 --> 00:14:06,560
societal

321
00:14:04,440 --> 00:14:08,759
expectations an interactive version of

322
00:14:06,560 --> 00:14:10,800
this project let users upload their own

323
00:14:08,759 --> 00:14:13,120
images to see how the AI chain

324
00:14:10,800 --> 00:14:14,600
reinterpreted them and that went about

325
00:14:13,120 --> 00:14:17,680
as well as you could

326
00:14:14,600 --> 00:14:23,079
expect many participants were surprised

327
00:14:17,680 --> 00:14:25,040
and um some were quite vifly upset uh by

328
00:14:23,079 --> 00:14:26,959
the results particularly when the system

329
00:14:25,040 --> 00:14:29,480
misgendered them uh the complaints I got

330
00:14:26,959 --> 00:14:33,040
were all from cisgender men white men I

331
00:14:29,480 --> 00:14:35,399
should say what a

332
00:14:33,040 --> 00:14:37,399
surprise uh this sparked conversations

333
00:14:35,399 --> 00:14:40,800
about how machines and people perceive

334
00:14:37,399 --> 00:14:43,440
identity and those those very real harms

335
00:14:40,800 --> 00:14:45,519
caused by

336
00:14:43,440 --> 00:14:47,639
assumptions this project unfortunately

337
00:14:45,519 --> 00:14:50,320
is no longer online because the systems

338
00:14:47,639 --> 00:14:52,880
that were being hosted um have been

339
00:14:50,320 --> 00:14:55,480
taken offline in favor of generative AI

340
00:14:52,880 --> 00:14:58,360
systems so I need to go back into the

341
00:14:55,480 --> 00:15:02,160
code rehost these um on a different

342
00:14:58,360 --> 00:15:04,639
platform set them up and um re

343
00:15:02,160 --> 00:15:07,600
reinitialize the project and I also want

344
00:15:04,639 --> 00:15:11,120
to take it into generative AI systems

345
00:15:07,600 --> 00:15:13,199
and redevelop it to see have have these

346
00:15:11,120 --> 00:15:18,320
systems evolved have they updated the

347
00:15:13,199 --> 00:15:20,680
images are better but are these biases

348
00:15:18,320 --> 00:15:23,040
in the chain do they still come out do

349
00:15:20,680 --> 00:15:26,120
the machines still misgender people do

350
00:15:23,040 --> 00:15:28,319
the systems still observe a new gender

351
00:15:26,120 --> 00:15:30,199
so that's what I'd love to contrast with

352
00:15:28,319 --> 00:15:32,360
the earlier versions see what

353
00:15:30,199 --> 00:15:34,519
developments if any have

354
00:15:32,360 --> 00:15:36,199
happened I've also put spun this

355
00:15:34,519 --> 00:15:38,759
Research into a series of image

356
00:15:36,199 --> 00:15:41,199
Generations across multiple stages of

357
00:15:38,759 --> 00:15:43,560
generative AI systems to see how they

358
00:15:41,199 --> 00:15:45,199
all interpret the same prompt this

359
00:15:43,560 --> 00:15:47,319
project is ongoing and may never really

360
00:15:45,199 --> 00:15:49,440
have an end date as new versions keep

361
00:15:47,319 --> 00:15:51,639
coming out but the viruses are very very

362
00:15:49,440 --> 00:15:54,600
clear I I should explain that my

363
00:15:51,639 --> 00:15:57,639
methodology here was that it was the

364
00:15:54,600 --> 00:15:59,079
middle of winter and I was very cold um

365
00:15:57,639 --> 00:16:01,480
and I think it was during lockdown as

366
00:15:59,079 --> 00:16:05,079
well so I wanted to be at the

367
00:16:01,480 --> 00:16:07,279
beach this is this is a legitimate

368
00:16:05,079 --> 00:16:09,759
methodology so just in looking a

369
00:16:07,279 --> 00:16:12,240
generative images of non-binary people

370
00:16:09,759 --> 00:16:14,240
I've noticed the same Trends every

371
00:16:12,240 --> 00:16:16,279
person is Young of course they're all

372
00:16:14,240 --> 00:16:19,240
gorgeous almost all of them have the

373
00:16:16,279 --> 00:16:21,920
short fairy hair pale skin and most are

374
00:16:19,240 --> 00:16:23,519
fem despite the many non-binary people I

375
00:16:21,920 --> 00:16:26,279
know with beards I've actually never

376
00:16:23,519 --> 00:16:29,399
seen any on a generative on a generated

377
00:16:26,279 --> 00:16:32,639
non-binary person without specifically

378
00:16:29,399 --> 00:16:34,399
asking actually apart from the age they

379
00:16:32,639 --> 00:16:36,959
all tend to look very

380
00:16:34,399 --> 00:16:40,519
similar which is great for all of the

381
00:16:36,959 --> 00:16:42,519
people who fit in that box I guess um

382
00:16:40,519 --> 00:16:45,360
but what about non-binary people of

383
00:16:42,519 --> 00:16:47,880
color what about high fem or high mask

384
00:16:45,360 --> 00:16:50,199
non-binary people it's so very hard to

385
00:16:47,880 --> 00:16:53,560
get a non-white non-binary person that

386
00:16:50,199 --> 00:16:56,319
even if I explicitly State an ethnicity

387
00:16:53,560 --> 00:16:58,360
it still struggles the eurocentric

388
00:16:56,319 --> 00:17:01,279
biases of these systems have always been

389
00:16:58,360 --> 00:17:05,639
really wrong but for non-binary

390
00:17:01,279 --> 00:17:05,639
representation it's an almost impassible

391
00:17:05,799 --> 00:17:10,880
barrier I'm taking similar directions in

392
00:17:08,720 --> 00:17:14,000
some of my new work to explore how image

393
00:17:10,880 --> 00:17:17,039
toex systems view a pixelated image of a

394
00:17:14,000 --> 00:17:19,039
person and reinterpret it as a new image

395
00:17:17,039 --> 00:17:22,319
this shows the biases of the system very

396
00:17:19,039 --> 00:17:24,520
clearly and by making it personal shows

397
00:17:22,319 --> 00:17:27,079
how these systems average out features

398
00:17:24,520 --> 00:17:29,600
and remove individuality through their

399
00:17:27,079 --> 00:17:31,120
biased data actually really like this

400
00:17:29,600 --> 00:17:33,799
image

401
00:17:31,120 --> 00:17:38,559
though I for I didn't save the caption

402
00:17:33,799 --> 00:17:41,400
though um uh I wasn't wild the captions

403
00:17:38,559 --> 00:17:43,080
are quite horrible they're through clip

404
00:17:41,400 --> 00:17:44,600
but I really actually quite liked that

405
00:17:43,080 --> 00:17:46,160
result I felt very

406
00:17:44,600 --> 00:17:49,520
validated

407
00:17:46,160 --> 00:17:53,320
um sometimes these systems don't

408
00:17:49,520 --> 00:17:53,320
suck no they do they all

409
00:17:53,440 --> 00:17:59,240
suck the final project in my PhD gender

410
00:17:56,760 --> 00:18:01,400
tapestry brought together the insights

411
00:17:59,240 --> 00:18:04,039
from my earlier Works into a unified

412
00:18:01,400 --> 00:18:07,120
piece this interactive project aimed to

413
00:18:04,039 --> 00:18:09,799
reframe G gender classification as a

414
00:18:07,120 --> 00:18:12,440
spectrum rather than as a binary using a

415
00:18:09,799 --> 00:18:15,880
custom classifier that assigned each

416
00:18:12,440 --> 00:18:18,159
participant a unique

417
00:18:15,880 --> 00:18:20,360
color the data set for this work was

418
00:18:18,159 --> 00:18:22,799
mostly generated it felt less

419
00:18:20,360 --> 00:18:25,440
problematic to use generated images of

420
00:18:22,799 --> 00:18:28,840
people and originally I had an even

421
00:18:25,440 --> 00:18:30,919
spread across genders however I found

422
00:18:28,840 --> 00:18:33,679
that the color results in that case were

423
00:18:30,919 --> 00:18:37,360
very very limited I actually needed to

424
00:18:33,679 --> 00:18:39,679
bias the results or bias the data set

425
00:18:37,360 --> 00:18:42,880
which really surprised me that was a

426
00:18:39,679 --> 00:18:44,480
crazy thing um I've been fighting bias

427
00:18:42,880 --> 00:18:46,559
my whole machine learning life and now

428
00:18:44,480 --> 00:18:49,559
you're telling me I need to put in bias

429
00:18:46,559 --> 00:18:50,679
but the data didn't lie on that front so

430
00:18:49,559 --> 00:18:53,760
I

431
00:18:50,679 --> 00:18:56,640
actually biased the data set in line

432
00:18:53,760 --> 00:18:59,440
with current representational data um I

433
00:18:56,640 --> 00:19:05,559
still kept it as um diverse as I

434
00:18:59,440 --> 00:19:11,080
possibly could but you know um and I I

435
00:19:05,559 --> 00:19:13,559
sorted them into six different pronoun

436
00:19:11,080 --> 00:19:15,240
categories users upload an image of

437
00:19:13,559 --> 00:19:17,440
their face which is analyzed by a

438
00:19:15,240 --> 00:19:19,799
multi-label classifier trained on these

439
00:19:17,440 --> 00:19:21,960
pronoun based data sets instead of

440
00:19:19,799 --> 00:19:25,159
assigning a gender the classifier

441
00:19:21,960 --> 00:19:30,360
outputs percentages that are then uh

442
00:19:25,159 --> 00:19:33,720
added together to create a an R GB

443
00:19:30,360 --> 00:19:35,840
value these values were used to generate

444
00:19:33,720 --> 00:19:38,200
a custom color which represented the

445
00:19:35,840 --> 00:19:41,559
user's unique experience of gender what

446
00:19:38,200 --> 00:19:44,679
I particularly enjoyed here was that um

447
00:19:41,559 --> 00:19:46,720
by taking six and turning it into three

448
00:19:44,679 --> 00:19:49,240
by adding those those disparate Parts

449
00:19:46,720 --> 00:19:51,799
together I created some Randomness and

450
00:19:49,240 --> 00:19:56,080
then if the number went over

451
00:19:51,799 --> 00:19:57,440
255 the modulo operator kicked in and um

452
00:19:56,080 --> 00:19:59,600
brought it brought the number back

453
00:19:57,440 --> 00:20:02,000
around which again increased that

454
00:19:59,600 --> 00:20:03,600
Randomness so while these

455
00:20:02,000 --> 00:20:08,039
classifications and these colors are

456
00:20:03,600 --> 00:20:10,679
based on fact they're actually all still

457
00:20:08,039 --> 00:20:12,919
um they're based on on a scientific

458
00:20:10,679 --> 00:20:14,480
methodology and on on proper

459
00:20:12,919 --> 00:20:17,039
classification but there're still an

460
00:20:14,480 --> 00:20:19,799
element of you know magic and weirdness

461
00:20:17,039 --> 00:20:21,520
and math that helps create those

462
00:20:19,799 --> 00:20:23,880
beautiful random

463
00:20:21,520 --> 00:20:25,840
effects you can actually uh it's

464
00:20:23,880 --> 00:20:28,440
probably um still up at gender

465
00:20:25,840 --> 00:20:30,720
tapestry.com uh the Gan won't be running

466
00:20:28,440 --> 00:20:32,679
but the rest is the users color and

467
00:20:30,720 --> 00:20:34,840
portrait were then added to a growing

468
00:20:32,679 --> 00:20:38,720
data set which was used to generate a

469
00:20:34,840 --> 00:20:41,159
mosaic of can created faces over time

470
00:20:38,720 --> 00:20:43,520
the Mosaic became increasingly complex

471
00:20:41,159 --> 00:20:47,200
as more colors and images were added

472
00:20:43,520 --> 00:20:50,720
illustrating that diversity of gender

473
00:20:47,200 --> 00:20:53,640
experience the concept of a of assigning

474
00:20:50,720 --> 00:20:56,080
a color to gender stem from the idea

475
00:20:53,640 --> 00:20:58,240
that like color perception gender is

476
00:20:56,080 --> 00:21:00,440
subjective and deeply personal my

477
00:20:58,240 --> 00:21:03,720
research revealed that we all actually

478
00:21:00,440 --> 00:21:05,320
see color differently the blue sky that

479
00:21:03,720 --> 00:21:06,480
I see is very different from the Blue

480
00:21:05,320 --> 00:21:10,320
Sky you

481
00:21:06,480 --> 00:21:13,279
see so we it's actually all a very very

482
00:21:10,320 --> 00:21:15,840
subjective experience and gender is a

483
00:21:13,279 --> 00:21:17,440
subjective experience and it's all

484
00:21:15,840 --> 00:21:19,679
deeply personal so while two people

485
00:21:17,440 --> 00:21:22,320
might have similar experiences no two

486
00:21:19,679 --> 00:21:24,159
are exactly alike this approach

487
00:21:22,320 --> 00:21:27,080
challenge the feasibility of gender

488
00:21:24,159 --> 00:21:29,520
classification by AI emphasizing that

489
00:21:27,080 --> 00:21:32,279
gender isn't something that can or

490
00:21:29,520 --> 00:21:34,880
should be reduced to a

491
00:21:32,279 --> 00:21:37,720
binary across these projects I explored

492
00:21:34,880 --> 00:21:39,760
the ways AI reflects human biases and

493
00:21:37,720 --> 00:21:42,600
the potential for Creative interventions

494
00:21:39,760 --> 00:21:46,360
to challenge those biases some of my

495
00:21:42,600 --> 00:21:49,440
findings include shocker a bias is

496
00:21:46,360 --> 00:21:51,480
deeply embedded in data sets who knew

497
00:21:49,440 --> 00:21:53,960
even when AI systems are trained on

498
00:21:51,480 --> 00:21:56,120
neutral data the assumptions of their

499
00:21:53,960 --> 00:21:58,840
creators can shape their outputs in the

500
00:21:56,120 --> 00:22:01,320
paper biased data sets are not enough

501
00:21:58,840 --> 00:22:04,159
results like uh a classifiers would

502
00:22:01,320 --> 00:22:07,200
classify an image uh as a woman if a

503
00:22:04,159 --> 00:22:09,960
person appeared in a kitchen because

504
00:22:07,200 --> 00:22:12,240
photographs of women were more common in

505
00:22:09,960 --> 00:22:15,200
kitchens than photographs of men in

506
00:22:12,240 --> 00:22:19,240
kitchens just a small example there of

507
00:22:15,200 --> 00:22:22,840
some yeah systemic bias there AI can

508
00:22:19,240 --> 00:22:25,279
evolve but not perfectly while new data

509
00:22:22,840 --> 00:22:27,559
can shift a model's perspective it often

510
00:22:25,279 --> 00:22:29,840
retains traces of its original biases

511
00:22:27,559 --> 00:22:32,919
much like humans do

512
00:22:29,840 --> 00:22:34,880
representation matters expanding data

513
00:22:32,919 --> 00:22:37,520
sets to include more diverse identities

514
00:22:34,880 --> 00:22:40,360
is crucial for creating more Equitable

515
00:22:37,520 --> 00:22:44,000
systems but it's only part of the

516
00:22:40,360 --> 00:22:46,520
solution e through my practice I hope to

517
00:22:44,000 --> 00:22:48,679
highlight not only the limitations of AI

518
00:22:46,520 --> 00:22:52,120
but also its potential as a medium for

519
00:22:48,679 --> 00:22:54,440
reflection critique and growth AI

520
00:22:52,120 --> 00:22:55,360
systems may never fully compute what

521
00:22:54,440 --> 00:22:58,520
gender

522
00:22:55,360 --> 00:23:01,640
means but by questioning their outputs

523
00:22:58,520 --> 00:23:04,880
and our own biases we can move toward a

524
00:23:01,640 --> 00:23:09,000
more inclusive future the best time to

525
00:23:04,880 --> 00:23:13,400
start is yesterday we know that but the

526
00:23:09,000 --> 00:23:16,880
next best time is now today so initially

527
00:23:13,400 --> 00:23:18,480
humans are told how to perceive gender

528
00:23:16,880 --> 00:23:21,440
they learn what gender is from their

529
00:23:18,480 --> 00:23:23,559
parents teachers and peers they grow and

530
00:23:21,440 --> 00:23:25,679
learn they learn to perceive gender how

531
00:23:23,559 --> 00:23:28,559
they want to perceive it they research

532
00:23:25,679 --> 00:23:30,400
or they don't they learn and listen and

533
00:23:28,559 --> 00:23:31,360
their understanding of the concept of

534
00:23:30,400 --> 00:23:34,440
gender

535
00:23:31,360 --> 00:23:37,360
grows similarly machines are taught how

536
00:23:34,440 --> 00:23:40,120
to perceive gender by humans they are

537
00:23:37,360 --> 00:23:43,200
shown images and told that these images

538
00:23:40,120 --> 00:23:45,679
correspond to genders an AI will not

539
00:23:43,200 --> 00:23:48,559
update its definitions on gender unless

540
00:23:45,679 --> 00:23:51,360
instructed to it will not change its

541
00:23:48,559 --> 00:23:54,480
perspective or enhance its understanding

542
00:23:51,360 --> 00:23:57,360
unless it is specifically retrained it

543
00:23:54,480 --> 00:23:59,080
can only do as its instructions permit

544
00:23:57,360 --> 00:24:02,159
it then falls to the people training the

545
00:23:59,080 --> 00:24:05,240
AI to and constructing the data sets to

546
00:24:02,159 --> 00:24:07,440
ensure that it's trained without bias a

547
00:24:05,240 --> 00:24:09,760
computer itself has no inherent biases

548
00:24:07,440 --> 00:24:11,840
it's only taught bias as it learns and

549
00:24:09,760 --> 00:24:14,080
incorporates bias data and

550
00:24:11,840 --> 00:24:16,559
algorithms the people working with the

551
00:24:14,080 --> 00:24:18,559
AI therefore need to understand the

552
00:24:16,559 --> 00:24:21,520
potential for bias and the inherent

553
00:24:18,559 --> 00:24:23,400
issues in facial classification tasks

554
00:24:21,520 --> 00:24:25,559
including gender classification

555
00:24:23,400 --> 00:24:27,919
especially gender

556
00:24:25,559 --> 00:24:31,080
classification people are naturally

557
00:24:27,919 --> 00:24:34,919
consciously or unconsciously biased and

558
00:24:31,080 --> 00:24:37,679
their inherent biases creep into their

559
00:24:34,919 --> 00:24:40,000
work we can help subvert bias by being

560
00:24:37,679 --> 00:24:42,760
self-aware of our own biases and our own

561
00:24:40,000 --> 00:24:45,480
issues gender is an Ever evolving social

562
00:24:42,760 --> 00:24:48,039
construct informed by experience and

563
00:24:45,480 --> 00:24:49,760
upbringing similarly AI is an Ever

564
00:24:48,039 --> 00:24:51,760
evolving machine construct with human

565
00:24:49,760 --> 00:24:54,440
supplied inputs that could be kind of

566
00:24:51,760 --> 00:24:57,039
equated to a form of experience and

567
00:24:54,440 --> 00:24:58,960
upbringing raising an equitable AI

568
00:24:57,039 --> 00:25:01,720
requires the voices of many

569
00:24:58,960 --> 00:25:04,159
and the intention to listen when issues

570
00:25:01,720 --> 00:25:07,640
are exposed as we are currently in the

571
00:25:04,159 --> 00:25:10,080
nent stages of understanding AI now is

572
00:25:07,640 --> 00:25:12,960
the best time to explore how they

573
00:25:10,080 --> 00:25:15,120
interrelate untangling the convergence

574
00:25:12,960 --> 00:25:18,840
of machine learning and social change is

575
00:25:15,120 --> 00:25:22,200
a key requirement right now especially

576
00:25:18,840 --> 00:25:24,279
as AI development is so rapid the time

577
00:25:22,200 --> 00:25:26,840
is Rip to merge these developments and

578
00:25:24,279 --> 00:25:30,480
to realize we cannot train AI in the

579
00:25:26,840 --> 00:25:32,880
outdated social Moray of the past but

580
00:25:30,480 --> 00:25:34,799
rather think ahead to a new stage of

581
00:25:32,880 --> 00:25:38,000
development and understanding of

582
00:25:34,799 --> 00:25:40,240
ourselves and our AI we need to be aware

583
00:25:38,000 --> 00:25:42,520
and critical of our biases examining the

584
00:25:40,240 --> 00:25:45,399
data sets we create and the methods we

585
00:25:42,520 --> 00:25:47,960
use to train AI this vigilance isn't

586
00:25:45,399 --> 00:25:50,039
just about fairness it's about creating

587
00:25:47,960 --> 00:25:53,240
systems that reflect our Humanity in its

588
00:25:50,039 --> 00:25:55,520
fullness not just a narrow biased sliver

589
00:25:53,240 --> 00:25:57,840
of it and that brings me to why this

590
00:25:55,520 --> 00:25:58,840
work matters it's not just about Ai and

591
00:25:57,840 --> 00:26:01,440
data

592
00:25:58,840 --> 00:26:03,520
it's about us it's about how we Define

593
00:26:01,440 --> 00:26:05,360
ourselves and each other it's about how

594
00:26:03,520 --> 00:26:07,640
we interact with technology and how

595
00:26:05,360 --> 00:26:10,520
those interactions shape our future if

596
00:26:07,640 --> 00:26:13,360
we allow AI to have learned from our

597
00:26:10,520 --> 00:26:15,399
worst habits our biases prejudices and

598
00:26:13,360 --> 00:26:17,600
assumptions it will replicate and

599
00:26:15,399 --> 00:26:19,880
reinforce them in ways we might not even

600
00:26:17,600 --> 00:26:23,039
notice until it's too

601
00:26:19,880 --> 00:26:26,039
late but if we teach AI to recognize the

602
00:26:23,039 --> 00:26:28,320
diversity and fluidity of human identity

603
00:26:26,039 --> 00:26:30,679
if we build systems that acknowledge and

604
00:26:28,320 --> 00:26:33,200
celebrate the complexity of gender we're

605
00:26:30,679 --> 00:26:36,039
not just making better AI we're creating

606
00:26:33,200 --> 00:26:39,200
a better reflection of a better world

607
00:26:36,039 --> 00:26:42,200
one that sees and values everyone thank

608
00:26:39,200 --> 00:26:42,200
you