1
00:00:00,480 --> 00:00:03,480
foreign

2
00:00:08,340 --> 00:00:13,139
good morning everyone please welcome Dr

3
00:00:10,380 --> 00:00:15,880
J Rosenbaum with gender tapestry new

4
00:00:13,139 --> 00:00:20,660
ways of looking at gender classification

5
00:00:15,880 --> 00:00:24,119
[Applause]

6
00:00:20,660 --> 00:00:26,880
thank you so much ah

7
00:00:24,119 --> 00:00:29,279
so hi everyone my name is Dr Jay

8
00:00:26,880 --> 00:00:32,220
Rosenbaum and I've earned a PhD studying

9
00:00:29,279 --> 00:00:35,820
AI perceptions of gender at rmit

10
00:00:32,220 --> 00:00:38,700
I'm also a sessional lecturer at rmit

11
00:00:35,820 --> 00:00:41,100
and teach design students about the

12
00:00:38,700 --> 00:00:43,079
creative possibilities of AI and how to

13
00:00:41,100 --> 00:00:45,719
view it through a critical lens

14
00:00:43,079 --> 00:00:47,520
my work in AI centers around the

15
00:00:45,719 --> 00:00:49,379
navigation and evolution of our

16
00:00:47,520 --> 00:00:52,079
understanding of gender

17
00:00:49,379 --> 00:00:54,000
while I talk today I'm going to invite

18
00:00:52,079 --> 00:00:56,820
you all to upload portraits of yourself

19
00:00:54,000 --> 00:00:59,039
to this system and what should grow and

20
00:00:56,820 --> 00:01:01,140
change as more people use it

21
00:00:59,039 --> 00:01:03,300
upload a few different images if you

22
00:01:01,140 --> 00:01:07,260
really want to test it out so yeah this

23
00:01:03,300 --> 00:01:09,900
is a live demo uh what could go wrong

24
00:01:07,260 --> 00:01:13,140
I'm fascinated about the nascent stages

25
00:01:09,900 --> 00:01:16,140
of AI and how our evolving understanding

26
00:01:13,140 --> 00:01:18,479
of gender is entangled in our unfolding

27
00:01:16,140 --> 00:01:20,820
understanding of AI

28
00:01:18,479 --> 00:01:24,060
also interested in the ethics of gender

29
00:01:20,820 --> 00:01:25,140
classification engendered images made by

30
00:01:24,060 --> 00:01:27,720
AI

31
00:01:25,140 --> 00:01:30,240
so that's gendertapestry.com

32
00:01:27,720 --> 00:01:33,360
and yeah

33
00:01:30,240 --> 00:01:36,540
perception is unique to the individual

34
00:01:33,360 --> 00:01:38,159
no person sees the same thing in quite

35
00:01:36,540 --> 00:01:40,799
the same way as another

36
00:01:38,159 --> 00:01:43,020
even color something we learn as

37
00:01:40,799 --> 00:01:45,600
children is perceived slightly

38
00:01:43,020 --> 00:01:48,000
differently by everyone

39
00:01:45,600 --> 00:01:50,820
not just by people with a perception

40
00:01:48,000 --> 00:01:54,680
impairment but evidence suggests that

41
00:01:50,820 --> 00:01:54,680
everyone perceives color differently

42
00:01:55,020 --> 00:02:02,820
a linear creation it's more accurately a

43
00:01:59,100 --> 00:02:06,479
three exists on a 3D color spectrum

44
00:02:02,820 --> 00:02:09,119
gender does not exist in a binary or as

45
00:02:06,479 --> 00:02:11,760
a linear Spectrum but in a 3D space

46
00:02:09,119 --> 00:02:13,440
that's highly subjective

47
00:02:11,760 --> 00:02:15,900
I want to discuss

48
00:02:13,440 --> 00:02:17,760
similarities between color perception

49
00:02:15,900 --> 00:02:19,800
and gender perception

50
00:02:17,760 --> 00:02:23,160
we all largely agree that the sky is

51
00:02:19,800 --> 00:02:24,780
blue right but what shade of blue

52
00:02:23,160 --> 00:02:27,420
exactly

53
00:02:24,780 --> 00:02:30,020
shade of blue that I

54
00:02:27,420 --> 00:02:30,020
Jade of blue

55
00:02:32,340 --> 00:02:37,020
be entirely different or gray as it is

56
00:02:34,980 --> 00:02:40,440
today

57
00:02:37,020 --> 00:02:43,500
is subjective and in the same way so is

58
00:02:40,440 --> 00:02:45,840
gender my experience of my non-binary

59
00:02:43,500 --> 00:02:47,819
gender is unique to me because it

60
00:02:45,840 --> 00:02:50,519
encompasses not only my socialized

61
00:02:47,819 --> 00:02:54,920
gender and my own relationship with my

62
00:02:50,519 --> 00:02:54,920
body but also the memories and

63
00:02:55,560 --> 00:03:00,180
cynic person that I am

64
00:02:57,540 --> 00:03:04,680
who I am

65
00:03:00,180 --> 00:03:07,620
experience of my gender is also unique

66
00:03:04,680 --> 00:03:09,840
impossible to fully equate one gender

67
00:03:07,620 --> 00:03:12,900
experience with another because our

68
00:03:09,840 --> 00:03:15,780
socialization our performativity and our

69
00:03:12,900 --> 00:03:18,780
bodies are all tied up inextricably in

70
00:03:15,780 --> 00:03:21,060
our gender and our gender expression

71
00:03:18,780 --> 00:03:24,599
gender people who have not explored

72
00:03:21,060 --> 00:03:27,000
their gender will still have a

73
00:03:24,599 --> 00:03:29,400
diff to other cisgender people because

74
00:03:27,000 --> 00:03:32,940
no one person experiences anything

75
00:03:29,400 --> 00:03:35,940
exactly the same way

76
00:03:32,940 --> 00:03:37,560
gender and color have a long history of

77
00:03:35,940 --> 00:03:40,680
being used in gendered marketing

78
00:03:37,560 --> 00:03:43,019
especially towards children and parents

79
00:03:40,680 --> 00:03:46,019
engender reveal parties baby clothes

80
00:03:43,019 --> 00:03:47,459
even the way we interpret colors and the

81
00:03:46,019 --> 00:03:49,319
names of colors

82
00:03:47,459 --> 00:03:52,860
although these colors were not always

83
00:03:49,319 --> 00:03:55,200
used this way today pink is synonymous

84
00:03:52,860 --> 00:03:57,360
with girls and blue with boys an

85
00:03:55,200 --> 00:03:58,980
epidemic of color and gender

86
00:03:57,360 --> 00:04:02,700
divisiveness

87
00:03:58,980 --> 00:04:05,099
we're told that these colors are blue or

88
00:04:02,700 --> 00:04:07,860
pink in the same way that we're told

89
00:04:05,099 --> 00:04:08,700
when growing up that we are a boy or a

90
00:04:07,860 --> 00:04:11,580
girl

91
00:04:08,700 --> 00:04:13,620
most people have little opportunity to

92
00:04:11,580 --> 00:04:16,019
challenge that idea

93
00:04:13,620 --> 00:04:18,720
but once we learn that there are other

94
00:04:16,019 --> 00:04:19,859
options and shades of difference between

95
00:04:18,720 --> 00:04:23,699
the two

96
00:04:19,859 --> 00:04:25,979
our experiences grow and we change our

97
00:04:23,699 --> 00:04:28,800
world widens

98
00:04:25,979 --> 00:04:33,800
this project is about widening that

99
00:04:28,800 --> 00:04:33,800
world about creating a tie Bing Bing

100
00:04:35,160 --> 00:04:38,300
no yes

101
00:04:41,720 --> 00:04:47,880
thank you

102
00:04:44,220 --> 00:04:50,479
sorry just thought I'd wait for that

103
00:04:47,880 --> 00:04:50,479
thank you

104
00:04:52,040 --> 00:04:56,699
this project is about widening that

105
00:04:54,360 --> 00:04:59,160
world about creating a tie between

106
00:04:56,699 --> 00:05:01,919
gender and color so that instead of

107
00:04:59,160 --> 00:05:05,340
receiving a gender classification people

108
00:05:01,919 --> 00:05:07,680
would receive their own unique color mix

109
00:05:05,340 --> 00:05:10,199
we often say gender is a spectrum

110
00:05:07,680 --> 00:05:11,880
without considering that it's more than

111
00:05:10,199 --> 00:05:13,259
a linear construct

112
00:05:11,880 --> 00:05:15,840
and so is color

113
00:05:13,259 --> 00:05:18,720
gender tapestry is a work that explores

114
00:05:15,840 --> 00:05:21,540
gender as a 3D color space

115
00:05:18,720 --> 00:05:24,360
the final result is an evolving Mosaic

116
00:05:21,540 --> 00:05:26,460
using the color results of the users and

117
00:05:24,360 --> 00:05:29,039
again trained on their faces

118
00:05:26,460 --> 00:05:31,620
the Mosaic gains in complexity as more

119
00:05:29,039 --> 00:05:35,340
people interact with the system and they

120
00:05:31,620 --> 00:05:38,460
receive their own custom gender colors

121
00:05:35,340 --> 00:05:40,740
enjoy working with multiple AI systems

122
00:05:38,460 --> 00:05:42,240
in a complex network of multimodal

123
00:05:40,740 --> 00:05:44,699
interaction

124
00:05:42,240 --> 00:05:46,259
throughout my research I've explored

125
00:05:44,699 --> 00:05:49,440
gender classification and its

126
00:05:46,259 --> 00:05:52,380
fallibility and lack of inclusiveness so

127
00:05:49,440 --> 00:05:54,120
I may have started gender tapestry with

128
00:05:52,380 --> 00:05:56,520
the aim of creating a better way to look

129
00:05:54,120 --> 00:05:59,340
at gender classification but in the

130
00:05:56,520 --> 00:06:01,860
process I think I've explored a bit of a

131
00:05:59,340 --> 00:06:04,199
better way to look at gender

132
00:06:01,860 --> 00:06:07,259
I recently showed gender tapestry at

133
00:06:04,199 --> 00:06:09,000
Acme in Melbourne and was delighted to

134
00:06:07,259 --> 00:06:11,160
see people's reactions as they found

135
00:06:09,000 --> 00:06:13,139
their faces in the mosaic

136
00:06:11,160 --> 00:06:15,840
it was also interesting to watch people

137
00:06:13,139 --> 00:06:18,000
agree to participate or not

138
00:06:15,840 --> 00:06:20,639
the majority of people who agreed to

139
00:06:18,000 --> 00:06:23,280
participate were younger with older men

140
00:06:20,639 --> 00:06:26,720
in particular running away at the mere

141
00:06:23,280 --> 00:06:26,720
mention of the word gender

142
00:06:27,900 --> 00:06:32,039
users interact with this work by

143
00:06:30,300 --> 00:06:34,500
uploading their image to the classifier

144
00:06:32,039 --> 00:06:36,539
it calculates a color result and

145
00:06:34,500 --> 00:06:37,740
presents the user with their image in

146
00:06:36,539 --> 00:06:39,960
that color

147
00:06:37,740 --> 00:06:42,300
they submit then becomes part of the

148
00:06:39,960 --> 00:06:44,039
training of the game and the color image

149
00:06:42,300 --> 00:06:46,500
that is generated becomes part of the

150
00:06:44,039 --> 00:06:50,240
Mosaic palette to reimagine the latest

151
00:06:46,500 --> 00:06:50,240
sample in the Gan generation

152
00:06:50,280 --> 00:06:53,240
gender classification

153
00:06:54,240 --> 00:06:58,500
gender classification relates entirely

154
00:06:56,639 --> 00:06:59,759
to the signs we teach computer vision

155
00:06:58,500 --> 00:07:02,460
agents

156
00:06:59,759 --> 00:07:05,340
most classifiers use an entirely binary

157
00:07:02,460 --> 00:07:09,440
jet system with all variables like hair

158
00:07:05,340 --> 00:07:09,440
makeup even glasses Stripped Away

159
00:07:09,539 --> 00:07:13,740
struggle classify people outside of the

160
00:07:12,180 --> 00:07:16,199
binary as well as passing and

161
00:07:13,740 --> 00:07:18,000
non-passing transgender people

162
00:07:16,199 --> 00:07:20,280
quit class of

163
00:07:18,000 --> 00:07:22,560
classify leaning into the expressive

164
00:07:20,280 --> 00:07:25,259
signals that people use to signify

165
00:07:22,560 --> 00:07:27,300
gender so that instead of teaching a

166
00:07:25,259 --> 00:07:30,180
classifier to just look at bone

167
00:07:27,300 --> 00:07:33,479
structure and eye distance and told it

168
00:07:30,180 --> 00:07:35,580
to see hair color and style makeup all

169
00:07:33,479 --> 00:07:37,440
the variable signs that go into the ways

170
00:07:35,580 --> 00:07:38,940
we express ourselves

171
00:07:37,440 --> 00:07:41,280
and our gender

172
00:07:38,940 --> 00:07:43,740
I wanted this to recognize gender as a

173
00:07:41,280 --> 00:07:46,560
performative construct rather than

174
00:07:43,740 --> 00:07:49,020
something built into our biology

175
00:07:46,560 --> 00:07:51,780
further create the classification by

176
00:07:49,020 --> 00:07:53,780
using a multi-label classifier that

177
00:07:51,780 --> 00:07:56,520
allows for more than one classification

178
00:07:53,780 --> 00:07:59,940
creating a set of percentages

179
00:07:56,520 --> 00:08:02,280
in this way I could just I could receive

180
00:07:59,940 --> 00:08:04,460
six sets of numbers to mix the colors

181
00:08:02,280 --> 00:08:04,460
with

182
00:08:04,500 --> 00:08:09,599
first stage in this project was to train

183
00:08:07,500 --> 00:08:12,720
the multi-label classifier in tensorflow

184
00:08:09,599 --> 00:08:14,220
or to put it more simply a non-binary

185
00:08:12,720 --> 00:08:16,560
classifier

186
00:08:14,220 --> 00:08:19,319
most image classifiers particularly

187
00:08:16,560 --> 00:08:22,139
gender classifiers operate on an either

188
00:08:19,319 --> 00:08:24,720
or system either the subject is female

189
00:08:22,139 --> 00:08:26,940
or male like comparing apples and

190
00:08:24,720 --> 00:08:29,520
oranges traditional classifiers choose

191
00:08:26,940 --> 00:08:31,259
between two things and if presented with

192
00:08:29,520 --> 00:08:34,260
something outside of those two choices

193
00:08:31,259 --> 00:08:36,300
they try to shoehorn it in to one or the

194
00:08:34,260 --> 00:08:38,520
other a classifier trained on either

195
00:08:36,300 --> 00:08:40,800
Barbie or Ken when presented with

196
00:08:38,520 --> 00:08:43,140
rollerblades we'll try to assign them to

197
00:08:40,800 --> 00:08:46,200
be either a Barbie or a kin

198
00:08:43,140 --> 00:08:47,940
it has no knowledge beyond the concept

199
00:08:46,200 --> 00:08:48,899
of the two choices it's been trained

200
00:08:47,940 --> 00:08:51,180
with

201
00:08:48,899 --> 00:08:53,820
a multi-label classifier on the other

202
00:08:51,180 --> 00:08:55,740
hand is trained with several labels and

203
00:08:53,820 --> 00:08:57,600
with images sometimes having more than

204
00:08:55,740 --> 00:08:59,940
one label attached

205
00:08:57,600 --> 00:09:02,339
this is the key when looking at gender

206
00:08:59,940 --> 00:09:05,580
as sometimes people have more than one

207
00:09:02,339 --> 00:09:08,279
label they like to go by thus instead of

208
00:09:05,580 --> 00:09:10,560
actually classifying gender I decided it

209
00:09:08,279 --> 00:09:14,600
would be less problematic to instead

210
00:09:10,560 --> 00:09:14,600
train the classifier in pronouns

211
00:09:14,760 --> 00:09:19,920
trans academic Dr Susan Striker

212
00:09:17,580 --> 00:09:22,500
discusses gender identity as something

213
00:09:19,920 --> 00:09:24,899
that could best be described as how one

214
00:09:22,500 --> 00:09:26,880
feels about being referred to by a

215
00:09:24,899 --> 00:09:29,220
particular pronoun

216
00:09:26,880 --> 00:09:31,380
the multi-label classifier allows me to

217
00:09:29,220 --> 00:09:33,959
train based on multiple pronoun options

218
00:09:31,380 --> 00:09:36,420
for the many non-binary people who use

219
00:09:33,959 --> 00:09:38,760
more than one set of pronouns

220
00:09:36,420 --> 00:09:41,339
the output of a multi-label classifier

221
00:09:38,760 --> 00:09:42,839
is a set of predictions presented as a

222
00:09:41,339 --> 00:09:45,420
percentage

223
00:09:42,839 --> 00:09:47,580
I worked with six pronoun labels in this

224
00:09:45,420 --> 00:09:50,940
classifier as six created more

225
00:09:47,580 --> 00:09:54,779
interesting variations in color

226
00:09:50,940 --> 00:09:57,300
I I developed the test classifier on

227
00:09:54,779 --> 00:10:00,360
generated faces so it seems a bizarre

228
00:09:57,300 --> 00:10:02,220
Prospect to assign gender to people that

229
00:10:00,360 --> 00:10:04,800
don't exist but it's actually really

230
00:10:02,220 --> 00:10:07,620
interesting philosophical premise the

231
00:10:04,800 --> 00:10:09,959
idea that an external person is signing

232
00:10:07,620 --> 00:10:12,180
gender or a machine classifying gender

233
00:10:09,959 --> 00:10:13,440
to a person they don't know seems really

234
00:10:12,180 --> 00:10:15,540
presumptive

235
00:10:13,440 --> 00:10:19,680
assigning gender labels to a fabricated

236
00:10:15,540 --> 00:10:21,959
person feels less problematic somehow

237
00:10:19,680 --> 00:10:23,100
pronouns I assign them are as artificial

238
00:10:21,959 --> 00:10:25,380
as they are

239
00:10:23,100 --> 00:10:26,519
as much of its construct as gender

240
00:10:25,380 --> 00:10:28,680
itself

241
00:10:26,519 --> 00:10:29,700
there is an artificiality to the entire

242
00:10:28,680 --> 00:10:32,100
project

243
00:10:29,700 --> 00:10:35,220
and I think that's an intriguing premise

244
00:10:32,100 --> 00:10:37,140
to classify real humans using a system

245
00:10:35,220 --> 00:10:38,580
that's been trained on artificial

246
00:10:37,140 --> 00:10:40,320
Humanity

247
00:10:38,580 --> 00:10:42,360
I chose pronouns instead of genders

248
00:10:40,320 --> 00:10:44,339
picks that fit neatly with the nature of

249
00:10:42,360 --> 00:10:46,380
the multi-label classifier because

250
00:10:44,339 --> 00:10:47,700
people often use more than one set of

251
00:10:46,380 --> 00:10:49,260
pronouns

252
00:10:47,700 --> 00:10:51,540
feel like there was something

253
00:10:49,260 --> 00:10:53,760
attractively artificial in the use of

254
00:10:51,540 --> 00:10:56,700
generated people to explore a social

255
00:10:53,760 --> 00:10:59,160
construct of gender and obscene

256
00:10:56,700 --> 00:11:00,600
assignation of something that's

257
00:10:59,160 --> 00:11:03,060
associated

258
00:11:00,600 --> 00:11:05,760
color for gender is a concept that

259
00:11:03,060 --> 00:11:07,800
already exists so why not subvert that

260
00:11:05,760 --> 00:11:09,480
and challenge the assumptions people

261
00:11:07,800 --> 00:11:10,860
have about each other and about

262
00:11:09,480 --> 00:11:13,920
themselves

263
00:11:10,860 --> 00:11:16,320
why not turn it into an experience as

264
00:11:13,920 --> 00:11:18,540
gender is an experience why not make

265
00:11:16,320 --> 00:11:23,000
something that shows how beautiful

266
00:11:18,540 --> 00:11:23,000
varied gender diversity really is

267
00:11:23,100 --> 00:11:27,420
I also found myself questioning the

268
00:11:25,500 --> 00:11:28,740
nature of bias as I developed this

269
00:11:27,420 --> 00:11:31,200
classifier

270
00:11:28,740 --> 00:11:34,620
when the data set was evenly distributed

271
00:11:31,200 --> 00:11:35,820
the results became murky too hard to

272
00:11:34,620 --> 00:11:38,519
distinguish

273
00:11:35,820 --> 00:11:40,380
I needed more categories and uneven

274
00:11:38,519 --> 00:11:42,079
distribution across the categories

275
00:11:40,380 --> 00:11:44,820
something closer to population

276
00:11:42,079 --> 00:11:47,160
representation in order to get the

277
00:11:44,820 --> 00:11:50,519
diversity of colors they wanted I

278
00:11:47,160 --> 00:11:52,980
actually needed to build in bias to

279
00:11:50,519 --> 00:11:55,980
receive varied results this is was

280
00:11:52,980 --> 00:11:57,120
antithetical to all of my research and

281
00:11:55,980 --> 00:11:59,640
beliefs

282
00:11:57,120 --> 00:12:01,860
I eventually had to use a public gender

283
00:11:59,640 --> 00:12:02,880
data set in addition to my fabricated

284
00:12:01,860 --> 00:12:04,860
one

285
00:12:02,880 --> 00:12:08,660
to skew the data set away from my

286
00:12:04,860 --> 00:12:08,660
idealized and balanced data

287
00:12:11,040 --> 00:12:16,200
but because I am also very much biased

288
00:12:14,279 --> 00:12:17,339
and it's a good practice for when you

289
00:12:16,200 --> 00:12:19,380
don't know people

290
00:12:17,339 --> 00:12:21,899
they them has the highest distribution

291
00:12:19,380 --> 00:12:26,300
and as you can see in the right hand

292
00:12:21,899 --> 00:12:26,300
column moving from three to six

293
00:12:26,399 --> 00:12:32,040
categories biases the data a little more

294
00:12:29,880 --> 00:12:33,839
just a little made an enormous

295
00:12:32,040 --> 00:12:37,019
difference

296
00:12:33,839 --> 00:12:39,720
even if I didn't like doing it

297
00:12:37,019 --> 00:12:41,399
this helped make the colors more vibrant

298
00:12:39,720 --> 00:12:44,160
and variable

299
00:12:41,399 --> 00:12:46,920
in the end the results I receive are

300
00:12:44,160 --> 00:12:50,279
consistent across images one picture

301
00:12:46,920 --> 00:12:52,860
will always receive the same color

302
00:12:50,279 --> 00:12:54,839
but when testing on myself I found that

303
00:12:52,860 --> 00:12:58,079
I received a range of colors based on my

304
00:12:54,839 --> 00:13:00,120
presentation at the time all usually in

305
00:12:58,079 --> 00:13:02,220
the same color family but subtly

306
00:13:00,120 --> 00:13:05,600
different

307
00:13:02,220 --> 00:13:05,600
after training the classifier

308
00:13:06,959 --> 00:13:10,740
after training the classifier I turned

309
00:13:09,120 --> 00:13:12,000
my attention to programming the key

310
00:13:10,740 --> 00:13:14,519
elements

311
00:13:12,000 --> 00:13:15,899
the prediction tool color generation and

312
00:13:14,519 --> 00:13:18,000
image generation

313
00:13:15,899 --> 00:13:20,899
this prediction uses the trained

314
00:13:18,000 --> 00:13:20,899
classifier model

315
00:13:21,959 --> 00:13:26,760
some custom code to process the results

316
00:13:24,120 --> 00:13:29,339
from a standard numpy array and turn

317
00:13:26,760 --> 00:13:31,800
them to individual integers that can be

318
00:13:29,339 --> 00:13:33,839
rendered in an RGB format

319
00:13:31,800 --> 00:13:35,880
so I'm gonna break down my programming

320
00:13:33,839 --> 00:13:39,000
that runs behind the scenes and explain

321
00:13:35,880 --> 00:13:40,620
not only how it works but why uh but

322
00:13:39,000 --> 00:13:43,860
also my decisions along the way so

323
00:13:40,620 --> 00:13:45,720
remember I am actually an artist and I

324
00:13:43,860 --> 00:13:50,720
don't necessarily have the best code

325
00:13:45,720 --> 00:13:53,760
this was before Chachi VT

326
00:13:50,720 --> 00:13:55,920
in this section the trained model is

327
00:13:53,760 --> 00:13:58,079
loaded the classes pronounced in this

328
00:13:55,920 --> 00:13:59,519
case are assigned to an array and

329
00:13:58,079 --> 00:14:01,920
variables are created for the

330
00:13:59,519 --> 00:14:03,720
probability of each result and the rank

331
00:14:01,920 --> 00:14:06,480
of each result

332
00:14:03,720 --> 00:14:10,139
Dynamic variables a generally bad

333
00:14:06,480 --> 00:14:11,360
practice but in this case it was

334
00:14:10,139 --> 00:14:14,220
necessary

335
00:14:11,360 --> 00:14:16,920
as the number was output in a single

336
00:14:14,220 --> 00:14:19,139
numpy array and I needed to split that

337
00:14:16,920 --> 00:14:22,100
array up in order to work with the

338
00:14:19,139 --> 00:14:22,100
numbers individually

339
00:14:22,200 --> 00:14:27,480
this section evaluates the image through

340
00:14:24,839 --> 00:14:29,399
this through each of the six results and

341
00:14:27,480 --> 00:14:32,160
assigns a prediction for each pronoun

342
00:14:29,399 --> 00:14:34,380
category based on the trained model

343
00:14:32,160 --> 00:14:37,680
it initially outputs a series of numbers

344
00:14:34,380 --> 00:14:39,420
that are all under one so each number is

345
00:14:37,680 --> 00:14:41,760
multiplied by a hundred so that we deal

346
00:14:39,420 --> 00:14:44,459
with whole numbers ready for mapping

347
00:14:41,760 --> 00:14:46,380
the top six variable contains the format

348
00:14:44,459 --> 00:14:49,079
for defining the order in which the

349
00:14:46,380 --> 00:14:50,760
array is sorted and the for Loop ensures

350
00:14:49,079 --> 00:14:53,100
that the prediction runs through each

351
00:14:50,760 --> 00:14:54,959
format and assigns a dynamic variable to

352
00:14:53,100 --> 00:14:57,240
each

353
00:14:54,959 --> 00:14:59,820
this section separates out the proper

354
00:14:57,240 --> 00:15:02,399
variable into six variables and cleans

355
00:14:59,820 --> 00:15:04,620
them ready for processing it uses the

356
00:15:02,399 --> 00:15:06,720
find function to isolate each set of

357
00:15:04,620 --> 00:15:07,800
numbers and assign them a dynamic

358
00:15:06,720 --> 00:15:10,440
variable

359
00:15:07,800 --> 00:15:12,839
and then to a variable redundantly

360
00:15:10,440 --> 00:15:15,540
called hex code

361
00:15:12,839 --> 00:15:18,720
if originally I designed this project to

362
00:15:15,540 --> 00:15:21,480
work with hex codes and if I were to

363
00:15:18,720 --> 00:15:22,800
stick with those all I would need to do

364
00:15:21,480 --> 00:15:25,380
is add them together at this point

365
00:15:22,800 --> 00:15:28,620
however the hex card colors were

366
00:15:25,380 --> 00:15:31,380
invariably boring not vibrant like I

367
00:15:28,620 --> 00:15:34,199
wanted so I formatted it to work with an

368
00:15:31,380 --> 00:15:37,800
RGB code instead

369
00:15:34,199 --> 00:15:39,860
my concern with using RGB colors was

370
00:15:37,800 --> 00:15:45,000
that many results ended up being over

371
00:15:39,860 --> 00:15:47,220
255. the maximum number of an RGB color

372
00:15:45,000 --> 00:15:48,120
so I set it loose and I was expecting it

373
00:15:47,220 --> 00:15:50,339
to fail

374
00:15:48,120 --> 00:15:52,980
but received instead a plethora of

375
00:15:50,339 --> 00:15:55,260
colors exactly what I wanted

376
00:15:52,980 --> 00:15:57,420
I examined the colors in a spreadsheet

377
00:15:55,260 --> 00:16:00,240
trying to see if the higher numbers were

378
00:15:57,420 --> 00:16:03,540
chopped off at 255 or normalized in the

379
00:16:00,240 --> 00:16:06,420
middle at 128 but neither was correct

380
00:16:03,540 --> 00:16:10,980
inside the color processing card any

381
00:16:06,420 --> 00:16:13,440
number of a 255 was divided by 255 and

382
00:16:10,980 --> 00:16:16,380
the difference taken is the output

383
00:16:13,440 --> 00:16:19,019
so I tested this Theory by explicitly

384
00:16:16,380 --> 00:16:21,300
programming the modulo operator in as

385
00:16:19,019 --> 00:16:23,760
you can see and it made absolutely no

386
00:16:21,300 --> 00:16:27,300
changes to the output so I found the

387
00:16:23,760 --> 00:16:30,060
problem or the wonderful magic behind

388
00:16:27,300 --> 00:16:31,260
the the so I left it in because it

389
00:16:30,060 --> 00:16:33,839
always looks better to have it

390
00:16:31,260 --> 00:16:36,620
explicitly programmed in even if it was

391
00:16:33,839 --> 00:16:39,899
actually doing it behind the scenes

392
00:16:36,620 --> 00:16:42,600
the final step in the completed code was

393
00:16:39,899 --> 00:16:45,180
to render the uploaded image ironically

394
00:16:42,600 --> 00:16:47,699
as a binary image

395
00:16:45,180 --> 00:16:49,560
using the assigned colors to create two

396
00:16:47,699 --> 00:16:51,660
new versions of the uploaded photograph

397
00:16:49,560 --> 00:16:54,180
a light and a dark version

398
00:16:51,660 --> 00:16:57,440
in case the color was too light or too

399
00:16:54,180 --> 00:16:57,440
dark to register properly

400
00:16:58,199 --> 00:17:02,639
which libraries grayscale and

401
00:17:00,240 --> 00:17:04,500
thresholding abilities to generate a

402
00:17:02,639 --> 00:17:06,959
purely black and white image where I

403
00:17:04,500 --> 00:17:09,540
then specify the colors on the zero or

404
00:17:06,959 --> 00:17:10,799
one pixels and save it with a unique

405
00:17:09,540 --> 00:17:13,260
file name

406
00:17:10,799 --> 00:17:17,520
this turned it from an interesting

407
00:17:13,260 --> 00:17:19,980
experiment to something about the user

408
00:17:17,520 --> 00:17:22,439
from there the rest of the artwork

409
00:17:19,980 --> 00:17:24,839
started to take shape a way of

410
00:17:22,439 --> 00:17:28,160
visualizing this as a community-based

411
00:17:24,839 --> 00:17:28,160
interactive artwork

412
00:17:28,199 --> 00:17:33,419
the Mosaic and the Gan development are

413
00:17:31,020 --> 00:17:35,760
part of the interactiveness and as

414
00:17:33,419 --> 00:17:37,860
people use the work more and train the

415
00:17:35,760 --> 00:17:40,679
Gan further and add more colors to the

416
00:17:37,860 --> 00:17:43,140
Mosaic the more complex it grows

417
00:17:40,679 --> 00:17:46,020
this represents not only the spectrum of

418
00:17:43,140 --> 00:17:48,480
color but also our understanding of

419
00:17:46,020 --> 00:17:51,660
gender as it moves from very simple to

420
00:17:48,480 --> 00:17:54,780
increasingly complex beautiful in its

421
00:17:51,660 --> 00:17:58,080
variety and complexity

422
00:17:54,780 --> 00:18:00,660
as this work was completed in 2021 we

423
00:17:58,080 --> 00:18:02,700
were in lockdown and I was very aware

424
00:18:00,660 --> 00:18:05,100
that I may have to Pivot my entire PhD

425
00:18:02,700 --> 00:18:07,860
exhibition online

426
00:18:05,100 --> 00:18:10,559
so this work originally designed to be a

427
00:18:07,860 --> 00:18:13,400
site-specific projection work the became

428
00:18:10,559 --> 00:18:13,400
a web-based piece

429
00:18:14,160 --> 00:18:18,539
because I didn't want to rewrite

430
00:18:15,600 --> 00:18:20,400
everything again in JavaScript and

431
00:18:18,539 --> 00:18:23,400
tensorflow.js which I could have done

432
00:18:20,400 --> 00:18:24,720
but you know I I just didn't want to

433
00:18:23,400 --> 00:18:27,120
reinvent the wheel

434
00:18:24,720 --> 00:18:28,919
so our research ways to get python

435
00:18:27,120 --> 00:18:30,960
projects online and was reminded of

436
00:18:28,919 --> 00:18:33,240
Django and flask

437
00:18:30,960 --> 00:18:35,400
so I started learning everything I could

438
00:18:33,240 --> 00:18:39,120
about flask in order to put this project

439
00:18:35,400 --> 00:18:40,380
online and I had working relatively well

440
00:18:39,120 --> 00:18:42,240
but then I had

441
00:18:40,380 --> 00:18:46,380
I had upload service that over

442
00:18:42,240 --> 00:18:49,679
complicated everything being online I

443
00:18:46,380 --> 00:18:51,840
wanted it to be locked down to faces

444
00:18:49,679 --> 00:18:56,000
just faces

445
00:18:51,840 --> 00:18:56,000
no other body parts thank you

446
00:18:56,160 --> 00:19:01,860
so I got the basic site working but I

447
00:18:59,940 --> 00:19:04,440
struggled with hosting and connecting it

448
00:19:01,860 --> 00:19:07,260
to separate with the separate Gan server

449
00:19:04,440 --> 00:19:09,179
and with my deadline looming I Enlisted

450
00:19:07,260 --> 00:19:11,340
the aid of a friend to help me get it

451
00:19:09,179 --> 00:19:13,980
over the finish line and now it's hosted

452
00:19:11,340 --> 00:19:15,539
on flyer and works perfectly and that's

453
00:19:13,980 --> 00:19:18,140
the site that hopefully you're all

454
00:19:15,539 --> 00:19:18,140
testing out

455
00:19:20,460 --> 00:19:24,480
how to perceive

456
00:19:22,440 --> 00:19:27,960
perceive they learn what gender is from

457
00:19:24,480 --> 00:19:30,419
their parents teachers and peers and as

458
00:19:27,960 --> 00:19:32,520
they grow and learn they learn to

459
00:19:30,419 --> 00:19:35,880
perceive gender how they want to

460
00:19:32,520 --> 00:19:38,039
perceive it they research or they don't

461
00:19:35,880 --> 00:19:40,559
they learn and listen and their

462
00:19:38,039 --> 00:19:42,440
understanding of the concept of gender

463
00:19:40,559 --> 00:19:45,299
grows

464
00:19:42,440 --> 00:19:47,340
likewise machines are taught how to

465
00:19:45,299 --> 00:19:49,799
perceive gender by humans

466
00:19:47,340 --> 00:19:52,140
they're shown images and told that these

467
00:19:49,799 --> 00:19:55,679
images correspond to genders

468
00:19:52,140 --> 00:19:57,960
an AI will not update its definitions on

469
00:19:55,679 --> 00:20:00,419
gender unless instructed to

470
00:19:57,960 --> 00:20:02,880
it will not change its perspective or

471
00:20:00,419 --> 00:20:05,820
enhance its understanding unless it

472
00:20:02,880 --> 00:20:09,360
specifically retrained

473
00:20:05,820 --> 00:20:12,600
only do as its instructions permit

474
00:20:09,360 --> 00:20:15,000
then falls to the people training the AI

475
00:20:12,600 --> 00:20:18,240
and constructing the data sets to ensure

476
00:20:15,000 --> 00:20:20,460
that it's trained without bias

477
00:20:18,240 --> 00:20:23,100
a computer itself has no inherent bias

478
00:20:20,460 --> 00:20:26,039
it's only taught by us as it learns and

479
00:20:23,100 --> 00:20:28,980
incorporates biased data and algorithms

480
00:20:26,039 --> 00:20:30,600
the people working with the AI therefore

481
00:20:28,980 --> 00:20:32,940
need to understand the potential for

482
00:20:30,600 --> 00:20:36,000
bias and the inherent issues in facial

483
00:20:32,940 --> 00:20:37,500
classification tasks including gender

484
00:20:36,000 --> 00:20:39,840
classification

485
00:20:37,500 --> 00:20:41,940
people are naturally consciously or

486
00:20:39,840 --> 00:20:44,220
unconsciously biased and their inherent

487
00:20:41,940 --> 00:20:46,020
biases creep into their work as seen

488
00:20:44,220 --> 00:20:49,140
with mine

489
00:20:46,020 --> 00:20:51,960
we can help subvert bias in AI by being

490
00:20:49,140 --> 00:20:54,419
self-aware of our own biases and our own

491
00:20:51,960 --> 00:20:57,360
issues we can subvert bias by creating

492
00:20:54,419 --> 00:20:59,340
bias in other ways as I found out to

493
00:20:57,360 --> 00:21:02,880
make things more Equitable or more

494
00:20:59,340 --> 00:21:05,520
beautiful more diverse more interesting

495
00:21:02,880 --> 00:21:08,400
gender is an ever-evolving social

496
00:21:05,520 --> 00:21:11,460
construct informed by experience and

497
00:21:08,400 --> 00:21:13,980
upbringing likewise AI is an

498
00:21:11,460 --> 00:21:16,260
ever-evolving machine construct with

499
00:21:13,980 --> 00:21:18,900
human supplied inputs that could be

500
00:21:16,260 --> 00:21:20,100
equated to a form of experience and

501
00:21:18,900 --> 00:21:22,740
upbringing

502
00:21:20,100 --> 00:21:25,140
raising an equitable AI requires the

503
00:21:22,740 --> 00:21:28,020
voices of many and the intention to

504
00:21:25,140 --> 00:21:29,700
listen when issues are exposed as we're

505
00:21:28,020 --> 00:21:33,419
currently in the nascent stages of

506
00:21:29,700 --> 00:21:36,020
understanding AI now is the best

507
00:21:33,419 --> 00:21:36,020
explore

508
00:21:36,240 --> 00:21:40,440
explore untangling the convergence of

509
00:21:38,580 --> 00:21:43,559
machine learning and social change is a

510
00:21:40,440 --> 00:21:46,500
key requirement right now as especially

511
00:21:43,559 --> 00:21:49,140
as AI development is so rapid

512
00:21:46,500 --> 00:21:51,539
by the time my exegesis was published

513
00:21:49,140 --> 00:21:53,580
many of the Technologies discussed were

514
00:21:51,539 --> 00:21:57,720
considered obsolete

515
00:21:53,580 --> 00:22:00,059
however the themes and Concepts the

516
00:21:57,720 --> 00:22:03,120
questioning these Technologies and the

517
00:22:00,059 --> 00:22:05,100
exploration of gender must continue to

518
00:22:03,120 --> 00:22:07,620
be discussed

519
00:22:05,100 --> 00:22:09,720
during the term of my PhD some of the

520
00:22:07,620 --> 00:22:11,940
language around gender changed as we

521
00:22:09,720 --> 00:22:13,080
became more aware of each other and our

522
00:22:11,940 --> 00:22:16,380
identities

523
00:22:13,080 --> 00:22:18,539
as Ai and gender continue to evolve they

524
00:22:16,380 --> 00:22:20,400
must be considered together

525
00:22:18,539 --> 00:22:22,860
and informed by each other's

526
00:22:20,400 --> 00:22:25,500
understanding and development

527
00:22:22,860 --> 00:22:27,600
my work is about creating the nature

528
00:22:25,500 --> 00:22:29,400
it's about exploring the nature of

529
00:22:27,600 --> 00:22:30,720
gender and the nature of perceiving

530
00:22:29,400 --> 00:22:33,179
gender

531
00:22:30,720 --> 00:22:35,760
I've explored how to do bias gendered

532
00:22:33,179 --> 00:22:37,740
Works how AI create some genders images

533
00:22:35,760 --> 00:22:40,860
and how to create a new way of

534
00:22:37,740 --> 00:22:43,500
perceiving gender as a color spectrum

535
00:22:40,860 --> 00:22:45,179
we're at a critical juncture of

536
00:22:43,500 --> 00:22:48,179
technological and sociological

537
00:22:45,179 --> 00:22:50,820
advancement both with the surge in AI

538
00:22:48,179 --> 00:22:52,740
training and with recent sociological

539
00:22:50,820 --> 00:22:54,360
advancements in how we think and talk

540
00:22:52,740 --> 00:22:57,840
about gender

541
00:22:54,360 --> 00:23:00,299
time is ripe to merge those developments

542
00:22:57,840 --> 00:23:02,640
and realize that we can't train machines

543
00:23:00,299 --> 00:23:06,059
in the outdated social mores of the past

544
00:23:02,640 --> 00:23:08,159
but rather to think ahead to a new stage

545
00:23:06,059 --> 00:23:11,340
of development and understanding of

546
00:23:08,159 --> 00:23:14,120
ourselves and our AI

547
00:23:11,340 --> 00:23:14,120
thank you

548
00:23:19,799 --> 00:23:24,860
so much Dr J do we have any quick

549
00:23:22,500 --> 00:23:24,860
questions

550
00:23:30,720 --> 00:23:35,220
if someone was to come up to you and ask

551
00:23:32,760 --> 00:23:36,659
you what does my color mean how would

552
00:23:35,220 --> 00:23:39,240
you answer that question

553
00:23:36,659 --> 00:23:42,240
oh I like that question

554
00:23:39,240 --> 00:23:44,760
um so the fun part about it is because

555
00:23:42,240 --> 00:23:46,020
of that modulo operator confusing

556
00:23:44,760 --> 00:23:48,419
everything

557
00:23:46,020 --> 00:23:51,120
the colors don't map exactly so we don't

558
00:23:48,419 --> 00:23:53,640
have like you know red for she her and

559
00:23:51,120 --> 00:23:56,640
green for they them and blue for he him

560
00:23:53,640 --> 00:23:59,580
and so on we've got you know it's not

561
00:23:56,640 --> 00:24:02,520
it's not as cut and dried as that so the

562
00:23:59,580 --> 00:24:03,240
color means what you want to mean I

563
00:24:02,520 --> 00:24:05,480
think

564
00:24:03,240 --> 00:24:08,460
it means

565
00:24:05,480 --> 00:24:10,740
that if it's not what you expected and

566
00:24:08,460 --> 00:24:13,740
it probably isn't what one expects that

567
00:24:10,740 --> 00:24:17,120
maybe it's something to explore what

568
00:24:13,740 --> 00:24:20,659
does that color mean to you

569
00:24:17,120 --> 00:24:20,659
awesome thank you

570
00:24:24,000 --> 00:24:27,380
any other questions from anyone

571
00:24:37,200 --> 00:24:42,780
thank you for the presentation I really

572
00:24:39,900 --> 00:24:44,880
loved how you combined colors and biases

573
00:24:42,780 --> 00:24:46,799
I was just wondering how

574
00:24:44,880 --> 00:24:49,500
um this project could also be used to

575
00:24:46,799 --> 00:24:52,140
explore it onto other kinds of biases or

576
00:24:49,500 --> 00:24:54,360
kind of looking into how AI could work

577
00:24:52,140 --> 00:24:55,799
with other kinds of biases just as with

578
00:24:54,360 --> 00:24:59,220
gender bias

579
00:24:55,799 --> 00:25:00,960
I could probably use it in

580
00:24:59,220 --> 00:25:03,240
um different ways I actually use the

581
00:25:00,960 --> 00:25:05,400
same data set to create a different

582
00:25:03,240 --> 00:25:07,679
artwork at the science Gallery in

583
00:25:05,400 --> 00:25:09,780
Melbourne just recently where it

584
00:25:07,679 --> 00:25:12,419
assigned different personality trait

585
00:25:09,780 --> 00:25:15,240
labels uh to people

586
00:25:12,419 --> 00:25:18,299
based and it was it was largely a gender

587
00:25:15,240 --> 00:25:21,000
classifier but it was yeah it kind of it

588
00:25:18,299 --> 00:25:23,460
was exploring the gendered assumptions

589
00:25:21,000 --> 00:25:26,520
we have around personality traits

590
00:25:23,460 --> 00:25:28,140
assertive dominant nurturing those kinds

591
00:25:26,520 --> 00:25:32,760
of words

592
00:25:28,140 --> 00:25:36,000
um and then from there it was put into a

593
00:25:32,760 --> 00:25:38,400
stable diffusion output and then used in

594
00:25:36,000 --> 00:25:41,159
augmented reality but people could then

595
00:25:38,400 --> 00:25:43,500
select their own terms so that it worked

596
00:25:41,159 --> 00:25:45,480
on their terms not on the AIS which I

597
00:25:43,500 --> 00:25:48,539
thought was very important addition

598
00:25:45,480 --> 00:25:51,299
so I think that you could apply same

599
00:25:48,539 --> 00:25:54,240
classification kind of concept to a

600
00:25:51,299 --> 00:25:57,000
range of different ideas this is

601
00:25:54,240 --> 00:25:58,020
foremost a it's a art project uh it's

602
00:25:57,000 --> 00:26:00,480
very much

603
00:25:58,020 --> 00:26:03,900
um a night installation but yeah there's

604
00:26:00,480 --> 00:26:06,419
more things that you can do to

605
00:26:03,900 --> 00:26:08,580
you know add add to different biases and

606
00:26:06,419 --> 00:26:10,140
explore different options especially

607
00:26:08,580 --> 00:26:12,900
with the different weights and biases

608
00:26:10,140 --> 00:26:16,580
and the weighting of the data

609
00:26:12,900 --> 00:26:16,580
which I'm still examining

610
00:26:19,200 --> 00:26:23,520
thank you so much everyone for coming

611
00:26:21,240 --> 00:26:27,380
and thank you Jay Rosenberg we have your

612
00:26:23,520 --> 00:26:27,380
gift thank you so much

613
00:26:30,779 --> 00:26:34,679
we'll be back in 10 minutes with open

614
00:26:32,820 --> 00:26:38,279
source communities and 10 years of micro

615
00:26:34,679 --> 00:26:41,240
python with Damien George thank you

616
00:26:38,279 --> 00:26:41,240
thank you all so much