1
00:00:00,420 --> 00:00:05,910
[Music]

2
00:00:10,719 --> 00:00:17,720
Welcome to the education track, the

3
00:00:14,799 --> 00:00:21,780
student showcase. Yeah,

4
00:00:17,720 --> 00:00:21,780
[Applause]

5
00:00:22,400 --> 00:00:28,279
this is my favorite time of year where I

6
00:00:24,480 --> 00:00:28,279
have hope for the future.

7
00:00:29,119 --> 00:00:34,480
Uh we we love the student showcase. You

8
00:00:32,640 --> 00:00:36,559
love the student showcase. If you don't

9
00:00:34,480 --> 00:00:40,079
love it yet, you haven't been here

10
00:00:36,559 --> 00:00:42,320
before. These are students like of the

11
00:00:40,079 --> 00:00:45,520
high school variety and they do cool

12
00:00:42,320 --> 00:00:47,680
stuff. That's pretty cool. Um but before

13
00:00:45,520 --> 00:00:51,719
we do that,

14
00:00:47,680 --> 00:00:51,719
we thought we'd play a video.

15
00:00:54,399 --> 00:01:00,600
Fantastic day. I've had a fantastic day.

16
00:00:56,480 --> 00:01:00,600
Uh, we have a little bit of

17
00:01:04,159 --> 00:01:08,400
But before we do that, we thought we'd

18
00:01:05,840 --> 00:01:09,680
have some technical challenges.

19
00:01:08,400 --> 00:01:11,840
We're okay. We're okay.

20
00:01:09,680 --> 00:01:13,520
Okay.

21
00:01:11,840 --> 00:01:15,920
Um, I hope you've had a fantastic day.

22
00:01:13,520 --> 00:01:17,119
I've had a fantastic day. Uh, we have a

23
00:01:15,920 --> 00:01:19,200
little bit of a special thing that

24
00:01:17,119 --> 00:01:21,520
happens in the education track. Well,

25
00:01:19,200 --> 00:01:23,840
has happened for the last two years. uh

26
00:01:21,520 --> 00:01:26,000
in the afternoon session we hand over to

27
00:01:23,840 --> 00:01:27,759
high school students and find out what

28
00:01:26,000 --> 00:01:29,600
they've been doing with Python or what

29
00:01:27,759 --> 00:01:32,079
they've learned about programming and so

30
00:01:29,600 --> 00:01:34,320
we have some fantastic eight groups nine

31
00:01:32,079 --> 00:01:37,680
students to show you today their their

32
00:01:34,320 --> 00:01:39,680
work um and talk about their ideas. Hi

33
00:01:37,680 --> 00:01:41,119
Pon, I'm Summer Saunders and I spoke at

34
00:01:39,680 --> 00:01:43,040
the student showcase last year about

35
00:01:41,119 --> 00:01:45,040
cyber security and the pornogi.

36
00:01:43,040 --> 00:01:47,040
Currently I'm a first year at UTS for

37
00:01:45,040 --> 00:01:49,600
cyber security as well.

38
00:01:47,040 --> 00:01:51,360
Hi, I'm Jun and I did a student showcase

39
00:01:49,600 --> 00:01:53,600
last year on my program to take scans of

40
00:01:51,360 --> 00:01:55,600
trains from side. Right now I'm studying

41
00:01:53,600 --> 00:01:58,320
engineering at uni alongside learning

42
00:01:55,600 --> 00:01:59,920
computers for fun. And um at the

43
00:01:58,320 --> 00:02:01,680
conference last year, I really enjoyed

44
00:01:59,920 --> 00:02:04,000
some of the talks I listened to. So

45
00:02:01,680 --> 00:02:06,000
there was a one talk about implementing

46
00:02:04,000 --> 00:02:08,640
classes in Python 1.4 which I found

47
00:02:06,000 --> 00:02:10,959
really interesting. Um and I also really

48
00:02:08,640 --> 00:02:12,959
like meeting people. So the people from

49
00:02:10,959 --> 00:02:14,400
my student showcase and some of the

50
00:02:12,959 --> 00:02:17,200
people from my school that came to

51
00:02:14,400 --> 00:02:19,200
conference um I really liked hanging out

52
00:02:17,200 --> 00:02:22,319
with them at the conference. So that was

53
00:02:19,200 --> 00:02:24,239
really fun. I also um unexpectedly

54
00:02:22,319 --> 00:02:26,720
actually enjoyed like preparing for the

55
00:02:24,239 --> 00:02:28,720
talk a lot. So I found that really

56
00:02:26,720 --> 00:02:30,959
creative and really fun. So thank you so

57
00:02:28,720 --> 00:02:33,680
much for the opportunity and h um and

58
00:02:30,959 --> 00:02:36,720
happy 10th anniversary for PyCon. Hi

59
00:02:33,680 --> 00:02:38,640
PyCon, I'm Annabelle and in 2024 I

60
00:02:36,720 --> 00:02:40,720
presented my project Pythagoras at the

61
00:02:38,640 --> 00:02:42,160
student showcase. Uh Pythagoras is a

62
00:02:40,720 --> 00:02:44,480
small Python tool which you can use to

63
00:02:42,160 --> 00:02:46,879
generate SVGs that are themed from

64
00:02:44,480 --> 00:02:48,879
images. Uh anyways, now I currently

65
00:02:46,879 --> 00:02:51,360
volunteer at a nonprofit in the United

66
00:02:48,879 --> 00:02:53,599
States helping teenagers from all around

67
00:02:51,360 --> 00:02:55,680
the world build cool technical projects.

68
00:02:53,599 --> 00:02:57,680
Um yeah, I'm super grateful to have been

69
00:02:55,680 --> 00:02:59,280
able to attend PyCon and to show off my

70
00:02:57,680 --> 00:03:01,440
project as well as to see what other

71
00:02:59,280 --> 00:03:03,280
people in the Python uh Python community

72
00:03:01,440 --> 00:03:04,800
have been building. So, thanks for the

73
00:03:03,280 --> 00:03:06,959
opportunity.

74
00:03:04,800 --> 00:03:09,040
Hi, I'm Zach and I presented at PyCon

75
00:03:06,959 --> 00:03:10,480
2024 in the student showcase. I

76
00:03:09,040 --> 00:03:12,239
presented my Rocket League replay

77
00:03:10,480 --> 00:03:13,519
poster, which was a bot that spectates

78
00:03:12,239 --> 00:03:15,599
professional Rocket League player

79
00:03:13,519 --> 00:03:17,760
replays and posts them onto YouTube for

80
00:03:15,599 --> 00:03:20,159
everyone else to view. I am this year

81
00:03:17,760 --> 00:03:21,440
studying computer science at RMIT in the

82
00:03:20,159 --> 00:03:22,959
first year and I'm really looking

83
00:03:21,440 --> 00:03:25,120
forward to getting a job in tech or

84
00:03:22,959 --> 00:03:27,280
programming in the future. I really

85
00:03:25,120 --> 00:03:28,800
liked presenting at PyCon as I got to

86
00:03:27,280 --> 00:03:31,040
meet a bunch of cool people and present

87
00:03:28,800 --> 00:03:33,200
my project to everyone and it also made

88
00:03:31,040 --> 00:03:35,599
me refactor my code so it was more

89
00:03:33,200 --> 00:03:38,239
presentable to everyone. Thanks PyCon.

90
00:03:35,599 --> 00:03:41,200
Bye Pon. Happy 10 years of the education

91
00:03:38,239 --> 00:03:43,200
track. I'm Biana. I was in 2023 and I

92
00:03:41,200 --> 00:03:47,040
made a music recommendation system with

93
00:03:43,200 --> 00:03:50,480
the Spotify API. Um I loved PyCon because

94
00:03:47,040 --> 00:03:54,000
it started my sticker addiction. Yep.

95
00:03:50,480 --> 00:03:56,640
Um, right now I'm studying AI at UTS and

96
00:03:54,000 --> 00:03:59,519
I'm having a blast and I'm coaching a

97
00:03:56,640 --> 00:04:01,519
first robotics team. See you later.

98
00:03:59,519 --> 00:04:04,799
Hi, I'm Sean and for my education

99
00:04:01,519 --> 00:04:08,080
showcase, I demoed a script that I used

100
00:04:04,799 --> 00:04:10,080
to automate my attendance during co. Um,

101
00:04:08,080 --> 00:04:13,519
I'm currently studying CS and maths at

102
00:04:10,080 --> 00:04:15,680
UNSW and I enjoyed the showcase as it

103
00:04:13,519 --> 00:04:17,359
let me display this cool side project

104
00:04:15,680 --> 00:04:20,160
that I made.

105
00:04:17,359 --> 00:04:22,880
Hey, my name is Liv and for the 2021

106
00:04:20,160 --> 00:04:27,040
PyCon Education Showcase, I spoke about

107
00:04:22,880 --> 00:04:30,240
solar powered hearing aids. Um, and now

108
00:04:27,040 --> 00:04:32,400
I am studying engineering in the

109
00:04:30,240 --> 00:04:34,560
University of Technology Sydney along

110
00:04:32,400 --> 00:04:37,520
with a bachelor of arts in international

111
00:04:34,560 --> 00:04:40,400
studies. I'm currently in Zurich in

112
00:04:37,520 --> 00:04:42,880
Switzerland studying physics and I also

113
00:04:40,400 --> 00:04:47,280
work for a satellite company building

114
00:04:42,880 --> 00:04:49,120
satellites. Um, I really liked the

115
00:04:47,280 --> 00:04:51,440
student showcase because I got to see

116
00:04:49,120 --> 00:04:52,639
what everyone else was doing and

117
00:04:51,440 --> 00:04:56,080
everyone was really lovely and

118
00:04:52,639 --> 00:04:59,360
supportive. So, thanks to PyCon.

119
00:04:56,080 --> 00:05:01,840
Hi Pon, it's Manindra here. I was a

120
00:04:59,360 --> 00:05:04,320
student showcase three years ago at PyCon

121
00:05:01,840 --> 00:05:08,160
and I presented my project on uh

122
00:05:04,320 --> 00:05:10,560
debluring with deep learning. Um it was

123
00:05:08,160 --> 00:05:12,479
a group project which was taking digits

124
00:05:10,560 --> 00:05:14,479
and then we were basically blurring them

125
00:05:12,479 --> 00:05:16,720
and then debluring them using deep

126
00:05:14,479 --> 00:05:18,720
learning and my own neural network.

127
00:05:16,720 --> 00:05:20,560
Currently uh I'm a student at the

128
00:05:18,720 --> 00:05:24,639
Australian National University still

129
00:05:20,560 --> 00:05:26,560
doing more AI stuff. Um and yeah, ever

130
00:05:24,639 --> 00:05:29,520
since then I've been just loving AI and

131
00:05:26,560 --> 00:05:31,360
pursuing AI. Uh PyCon really helped build

132
00:05:29,520 --> 00:05:34,720
my confidence and text conferences

133
00:05:31,360 --> 00:05:38,240
weren't so scary. So I love PyCon for

134
00:05:34,720 --> 00:05:40,479
that. Anyways, bye guys. Bye PyCon.

135
00:05:38,240 --> 00:05:43,120
Hi everyone. My name's Angus. Back in

136
00:05:40,479 --> 00:05:45,280
2019, I participated in the PyCon AU

137
00:05:43,120 --> 00:05:47,600
education student showcase and it was

138
00:05:45,280 --> 00:05:48,960
really awesome. I got to share some of

139
00:05:47,600 --> 00:05:50,639
the adventures I've been having

140
00:05:48,960 --> 00:05:52,720
contributing to open source back in high

141
00:05:50,639 --> 00:05:54,320
school. And one of the things I really

142
00:05:52,720 --> 00:05:56,560
enjoyed about the conference was getting

143
00:05:54,320 --> 00:05:58,400
to meet a huge group of like-minded

144
00:05:56,560 --> 00:06:00,880
people who are really passionate about

145
00:05:58,400 --> 00:06:02,800
Python and about tech and sharing what

146
00:06:00,880 --> 00:06:05,600
they know with a community of awesome

147
00:06:02,800 --> 00:06:07,520
people. and I enjoyed it so much that I

148
00:06:05,600 --> 00:06:09,759
went on to study cyber security at uni.

149
00:06:07,520 --> 00:06:12,160
I now work in the field as a graduate

150
00:06:09,759 --> 00:06:14,000
and yeah, it's really awesome. I

151
00:06:12,160 --> 00:06:15,840
nowadays I contribute back to the

152
00:06:14,000 --> 00:06:18,240
community through by volunteering for

153
00:06:15,840 --> 00:06:20,400
our local hacker conference here in CRA

154
00:06:18,240 --> 00:06:22,319
and yeah, so I hope you guys really

155
00:06:20,400 --> 00:06:24,080
enjoy the conference. Give back what you

156
00:06:22,319 --> 00:06:26,000
can, help out. This is a really awesome

157
00:06:24,080 --> 00:06:28,000
community and everyone is willing to

158
00:06:26,000 --> 00:06:30,960
share so much and it's great.

159
00:06:28,000 --> 00:06:33,440
Hi, I'm Izzy. I'm 19. I'm in second year

160
00:06:30,960 --> 00:06:35,440
software engineering here in Sydney. I

161
00:06:33,440 --> 00:06:36,880
was part of the education track in 2018

162
00:06:35,440 --> 00:06:39,600
where I talked about my machine learning

163
00:06:36,880 --> 00:06:41,919
notes and crosses spot and 2023 where I

164
00:06:39,600 --> 00:06:44,560
talked about my year 12 major project. I

165
00:06:41,919 --> 00:06:46,000
then came back to PyCon last year as a

166
00:06:44,560 --> 00:06:48,160
volunteer where I did two lightning

167
00:06:46,000 --> 00:06:50,240
talks and I will be back again this year

168
00:06:48,160 --> 00:06:51,840
as the head of volunteers and I will be

169
00:06:50,240 --> 00:06:53,680
doing a full talk in the education

170
00:06:51,840 --> 00:06:55,120
stream. I just want to say thank you

171
00:06:53,680 --> 00:06:57,840
education track.

172
00:06:55,120 --> 00:07:02,080
Hi, I'm Sam. I presented at the student

173
00:06:57,840 --> 00:07:05,280
showcase in 2018, 2021, and 2022. I've

174
00:07:02,080 --> 00:07:07,039
been volunteering with AV for two years

175
00:07:05,280 --> 00:07:08,880
now, and this is going to be my fourth

176
00:07:07,039 --> 00:07:11,039
year presenting at PyCon. I'm doing a

177
00:07:08,880 --> 00:07:12,960
full talk this year. Uh I'll also be

178
00:07:11,039 --> 00:07:15,840
helping out with AV again this year. So,

179
00:07:12,960 --> 00:07:19,840
you might see me around. That's it.

180
00:07:15,840 --> 00:07:22,000
That's it. So, um we had a closing

181
00:07:19,840 --> 00:07:24,800
slide. The closing slide said thank you.

182
00:07:22,000 --> 00:07:29,319
Um, instead I will say thank you uh

183
00:07:24,800 --> 00:07:29,319
rather than risk uh more updates.

184
00:07:30,639 --> 00:07:36,880
What what was that technical challenges?

185
00:07:33,759 --> 00:07:39,360
I have no idea. Um I hope that gave you

186
00:07:36,880 --> 00:07:41,120
a little sense of how much the education

187
00:07:39,360 --> 00:07:42,319
track means to people who have gone

188
00:07:41,120 --> 00:07:44,479
through it in the community. We think

189
00:07:42,319 --> 00:07:47,440
it's really awesome. Um you're all here.

190
00:07:44,479 --> 00:07:50,560
You're stuck in the room anyway. Um so

191
00:07:47,440 --> 00:07:52,720
let's start. Our first speaker

192
00:07:50,560 --> 00:07:54,080
is not Ally.

193
00:07:52,720 --> 00:07:57,199
You'll be shocked to hear. Our first

194
00:07:54,080 --> 00:08:00,160
speaker is Arabella. Um she has been

195
00:07:57,199 --> 00:08:02,479
coding in Scratch since year three and

196
00:08:00,160 --> 00:08:06,879
in Python since oh sorry for seven

197
00:08:02,479 --> 00:08:11,080
years. Um her favorite emoji is

198
00:08:06,879 --> 00:08:11,080
definitely the skull emoji.

199
00:08:11,120 --> 00:08:17,919
I Yeah, sorry. Um uh something that she

200
00:08:15,360 --> 00:08:19,360
recently learned about computers. Um, I

201
00:08:17,919 --> 00:08:20,879
learned that when you access the cookies

202
00:08:19,360 --> 00:08:22,319
on a web page with JavaScript, they're

203
00:08:20,879 --> 00:08:24,160
not stored as a nicely formatted

204
00:08:22,319 --> 00:08:25,919
dictionary or anything, but literally as

205
00:08:24,160 --> 00:08:28,319
a big long string of names and values

206
00:08:25,919 --> 00:08:31,039
and are super super easy to set, but

207
00:08:28,319 --> 00:08:33,919
surprisingly hard to decode and get info

208
00:08:31,039 --> 00:08:40,839
from. Um, and she's going to be talking

209
00:08:33,919 --> 00:08:40,839
about the fact that phones aren't real.

210
00:08:43,039 --> 00:08:46,680
My name is Arabella Simpson and this is

211
00:08:45,600 --> 00:08:51,839
Phone

212
00:08:46,680 --> 00:08:53,920
[Music]

213
00:08:51,839 --> 00:08:56,399
A couple of years ago, I bought a 4G hat

214
00:08:53,920 --> 00:08:58,240
for my Raspberry Pi. Made a very simple

215
00:08:56,399 --> 00:09:00,320
SMS flask app to cure my phone

216
00:08:58,240 --> 00:09:02,480
addiction, which was lots of fun, but it

217
00:09:00,320 --> 00:09:04,959
was clunky and a bit useless. And then

218
00:09:02,480 --> 00:09:07,120
the hat broke. So last year, I decided

219
00:09:04,959 --> 00:09:09,279
to revamp the project into a smaller and

220
00:09:07,120 --> 00:09:10,560
more original form

221
00:09:09,279 --> 00:09:14,120
and incorporating one of my biggest

222
00:09:10,560 --> 00:09:14,120
passions, robotics.

223
00:09:14,160 --> 00:09:18,320
Now, the idea of a robot phone may seem

224
00:09:16,399 --> 00:09:20,320
shockingly creative to you, but it's an

225
00:09:18,320 --> 00:09:22,640
idea I've been considering for a while.

226
00:09:20,320 --> 00:09:25,120
Inspired by the YA novel Jinx by Amy

227
00:09:22,640 --> 00:09:29,000
Mulock and the satirical conspiracy

228
00:09:25,120 --> 00:09:29,000
theory Birds Aren't Real.

229
00:09:31,760 --> 00:09:35,920
I had seven requirements for this new

230
00:09:33,760 --> 00:09:38,800
project. Number one, it should have all

231
00:09:35,920 --> 00:09:41,600
the advantages of a modern mobile phone.

232
00:09:38,800 --> 00:09:44,160
Put that. Number two, should also have

233
00:09:41,600 --> 00:09:46,080
all the benefits of a robot. Number

234
00:09:44,160 --> 00:09:48,320
three, it should challenge the idea of a

235
00:09:46,080 --> 00:09:52,080
phone. And of course, it should be able

236
00:09:48,320 --> 00:09:56,279
to send text, take photos, record notes,

237
00:09:52,080 --> 00:09:56,279
and show the weather and time.

238
00:09:57,120 --> 00:10:01,760
To make my phone compact, I put all the

239
00:09:59,200 --> 00:10:03,600
Python code onto a Raspberry Pi Zero,

240
00:10:01,760 --> 00:10:05,680
and I designed and manufactured a custom

241
00:10:03,600 --> 00:10:07,279
PCB to mount all the electronic

242
00:10:05,680 --> 00:10:09,460
components and minimize the footprint

243
00:10:07,279 --> 00:10:13,029
and complexity of the design.

244
00:10:09,460 --> 00:10:13,029
[Music]

245
00:10:21,839 --> 00:10:27,440
To power my phone, I use 2,000 mAh

246
00:10:25,200 --> 00:10:29,120
battery, which is charged and provides 5

247
00:10:27,440 --> 00:10:31,519
vol power to the Raspberry Pi with the

248
00:10:29,120 --> 00:10:33,040
Adafruit Power Boost 500 board. This

249
00:10:31,519 --> 00:10:34,560
board has a switch soldered onto it,

250
00:10:33,040 --> 00:10:36,560
which allows me to turn the board on and

251
00:10:34,560 --> 00:10:37,839
off to save power so that it can last as

252
00:10:36,560 --> 00:10:40,240
long as possible without being

253
00:10:37,839 --> 00:10:41,920
recharged. Finally, the touchcreen

254
00:10:40,240 --> 00:10:44,480
interface is through an adorable round

255
00:10:41,920 --> 00:10:46,160
touchcreen LCD from Wisher. The

256
00:10:44,480 --> 00:10:48,320
documentation for the screen has some

257
00:10:46,160 --> 00:10:50,160
sample code on it, which I modified into

258
00:10:48,320 --> 00:10:53,320
a pillbased display class to generate

259
00:10:50,160 --> 00:10:53,320
the output.

260
00:10:53,440 --> 00:10:57,600
Right. So the second requirement I had

261
00:10:56,079 --> 00:11:00,399
was that it should have all the benefits

262
00:10:57,600 --> 00:11:03,360
of a robot which includes the Raspberry

263
00:11:00,399 --> 00:11:07,600
Pi which powers everything Python code

264
00:11:03,360 --> 00:11:12,000
and I wanted it to balance and move and

265
00:11:07,600 --> 00:11:13,760
stuff. Um yeah but in hindsight the self

266
00:11:12,000 --> 00:11:15,279
balancing was not such a great idea

267
00:11:13,760 --> 00:11:20,160
because this just launched itself off

268
00:11:15,279 --> 00:11:22,480
the table and um broke the broke the

269
00:11:20,160 --> 00:11:25,360
stand there. Also, the fact that I

270
00:11:22,480 --> 00:11:28,480
didn't make the stand tall enough, so it

271
00:11:25,360 --> 00:11:31,279
balances like has to be over the edge,

272
00:11:28,480 --> 00:11:33,680
which makes it more likely to launch

273
00:11:31,279 --> 00:11:36,320
yourself off the table. But the

274
00:11:33,680 --> 00:11:39,279
accelerometer balances

275
00:11:36,320 --> 00:11:42,240
as the roll and the angle of the servo

276
00:11:39,279 --> 00:11:45,480
trying to adjust for that, which is

277
00:11:42,240 --> 00:11:45,480
pretty good.

278
00:11:47,279 --> 00:11:54,079
I had lots of prototypes. started with

279
00:11:49,760 --> 00:11:58,240
sketches and then I did a really dodgy

280
00:11:54,079 --> 00:12:03,200
cardboard and tape model and some CAD

281
00:11:58,240 --> 00:12:05,920
modeling solid works and blender and

282
00:12:03,200 --> 00:12:07,440
made one out of core flute. I was

283
00:12:05,920 --> 00:12:09,920
originally going to do a pigeon, but

284
00:12:07,440 --> 00:12:12,240
pigeons, they have really tiny heads,

285
00:12:09,920 --> 00:12:14,320
weirdly long, skinny necks, which was

286
00:12:12,240 --> 00:12:17,200
quite a hassle for trying to deal with

287
00:12:14,320 --> 00:12:20,680
all the cables and trying to fit this

288
00:12:17,200 --> 00:12:20,680
massive camera.

289
00:12:22,079 --> 00:12:27,920
This is the final prototype. It's kind

290
00:12:25,040 --> 00:12:31,200
of inspired by Galar, although I didn't

291
00:12:27,920 --> 00:12:32,720
paint it, so the coloring is more cocku.

292
00:12:31,200 --> 00:12:35,279
we'd have like a conversation with the

293
00:12:32,720 --> 00:12:37,279
bird instead of interacting with the

294
00:12:35,279 --> 00:12:40,800
interface. So, what I've done is I've

295
00:12:37,279 --> 00:12:43,760
got a little PZO buzzer and originally

296
00:12:40,800 --> 00:12:47,279
on the Raspberry Pi 3B Plus, I connect

297
00:12:43,760 --> 00:12:49,120
that with a 3.5 mm connector to the

298
00:12:47,279 --> 00:12:51,600
audio jack.

299
00:12:49,120 --> 00:12:55,200
I use the ESE module as audio output to

300
00:12:51,600 --> 00:12:56,880
have the bird say the names of the apps

301
00:12:55,200 --> 00:13:00,000
that you're entering. The Raspberry Pi

302
00:12:56,880 --> 00:13:01,920
3B Plus doesn't have a 3.5 mil connector

303
00:13:00,000 --> 00:13:05,279
because it is so small than trying to

304
00:13:01,920 --> 00:13:08,399
save space. So instead, I've recreated

305
00:13:05,279 --> 00:13:11,200
on the PCB the original resistor

306
00:13:08,399 --> 00:13:13,200
capacitor circuit for the Raspberry Pi

307
00:13:11,200 --> 00:13:15,519
and I've got audio output through the

308
00:13:13,200 --> 00:13:18,160
GPIO pins. The audio is a little bit

309
00:13:15,519 --> 00:13:19,760
tiny

310
00:13:18,160 --> 00:13:21,680
because the transistors that I used

311
00:13:19,760 --> 00:13:23,440
didn't get wired correctly. So, I've

312
00:13:21,680 --> 00:13:26,399
actually bypassed them on the current

313
00:13:23,440 --> 00:13:28,639
prototype, but with the next PCB, I'll

314
00:13:26,399 --> 00:13:31,519
hopefully fix that issue. Number four

315
00:13:28,639 --> 00:13:34,800
was texts. I wanted to send and receive

316
00:13:31,519 --> 00:13:37,839
messages. Previously, I'd used the

317
00:13:34,800 --> 00:13:40,160
Raspberry Pi specific 4G hack. It's

318
00:13:37,839 --> 00:13:42,720
quite useful because there's dedicated

319
00:13:40,160 --> 00:13:45,360
Python libraries and it uses 80

320
00:13:42,720 --> 00:13:47,360
commands, which is how real life modems

321
00:13:45,360 --> 00:13:48,880
and phones work.

322
00:13:47,360 --> 00:13:53,440
I wanted to see if I could get it

323
00:13:48,880 --> 00:13:56,399
working just using a Telra 4G USB modem.

324
00:13:53,440 --> 00:13:58,880
And when I plug this into my PC, there's

325
00:13:56,399 --> 00:14:00,480
like a website that Telra has made that

326
00:13:58,880 --> 00:14:03,839
you connect to and then you can send

327
00:14:00,480 --> 00:14:05,760
text on there. But doesn't work so well

328
00:14:03,839 --> 00:14:07,360
for the Raspberry Pi, especially the

329
00:14:05,760 --> 00:14:11,680
Raspberry Pi Zero. It doesn't have

330
00:14:07,360 --> 00:14:13,680
enough RAM to actually use the browser.

331
00:14:11,680 --> 00:14:16,639
at commands that didn't work because of

332
00:14:13,680 --> 00:14:19,360
the software which is very user friendly

333
00:14:16,639 --> 00:14:21,600
if you have a PC not so user friendly if

334
00:14:19,360 --> 00:14:23,839
you're trying to make your own phone

335
00:14:21,600 --> 00:14:26,720
discovered that you could theoretically

336
00:14:23,839 --> 00:14:28,560
access the browser interface using

337
00:14:26,720 --> 00:14:32,240
Selenium which is an awesome Python

338
00:14:28,560 --> 00:14:34,160
library that um opens the browser and

339
00:14:32,240 --> 00:14:36,399
you can

340
00:14:34,160 --> 00:14:39,040
read stuff as you would with the request

341
00:14:36,399 --> 00:14:42,000
module but you can actually interact in

342
00:14:39,040 --> 00:14:43,680
real time um so I tried using that and I

343
00:14:42,000 --> 00:14:45,839
think it would be possible maybe to get

344
00:14:43,680 --> 00:14:50,519
working but haven't had any success at

345
00:14:45,839 --> 00:14:50,519
all even on a different Raspberry Pi.

346
00:14:52,800 --> 00:14:57,920
My fifth requirement was the camera app

347
00:14:55,600 --> 00:15:00,320
taking photos, recording videos and also

348
00:14:57,920 --> 00:15:02,800
looking back at the photos. Go into the

349
00:15:00,320 --> 00:15:06,680
menu and there's a camera app. This is a

350
00:15:02,800 --> 00:15:06,680
Raspberry Pi camera V2.

351
00:15:08,370 --> 00:15:15,580
[Music]

352
00:15:15,839 --> 00:15:22,480
Okay. So, the notes app, which is both a

353
00:15:20,560 --> 00:15:25,360
useful function to have on a phone, but

354
00:15:22,480 --> 00:15:28,399
also is sort of a template for how I'm

355
00:15:25,360 --> 00:15:30,560
going to implement the SMS for text if I

356
00:15:28,399 --> 00:15:32,160
ever get that working. So, the notes

357
00:15:30,560 --> 00:15:35,279
app, you've got two ways that you can

358
00:15:32,160 --> 00:15:37,120
make notes. Number one is through this

359
00:15:35,279 --> 00:15:39,519
microphone. microphone is mounted in the

360
00:15:37,120 --> 00:15:43,440
eye. USB microphone. You can dictate

361
00:15:39,519 --> 00:15:46,240
notes with that. Um, and I've got is it

362
00:15:43,440 --> 00:15:50,519
Google speech to text? I don't know. Um,

363
00:15:46,240 --> 00:15:50,519
anyway, you can dictate notes.

364
00:15:52,000 --> 00:15:56,240
Hello, this is some text. Please

365
00:15:54,240 --> 00:15:57,920
transcribe it.

366
00:15:56,240 --> 00:16:00,959
Three hours later.

367
00:15:57,920 --> 00:16:03,279
Uh, yeah, close enough, I guess. And the

368
00:16:00,959 --> 00:16:06,000
other type of note takingaking is a draw

369
00:16:03,279 --> 00:16:09,639
function where you can write notes like

370
00:16:06,000 --> 00:16:09,639
on an Apple Watch.

371
00:16:18,399 --> 00:16:23,120
Number seven was showing the weather and

372
00:16:20,880 --> 00:16:25,199
time on the home screen without having

373
00:16:23,120 --> 00:16:28,000
to open the phone or other apps and

374
00:16:25,199 --> 00:16:30,480
getting distracted. Yeah, I just used a

375
00:16:28,000 --> 00:16:33,040
basic weather API for that and the

376
00:16:30,480 --> 00:16:35,040
daytime module. But that's all I've got

377
00:16:33,040 --> 00:16:37,490
time for in this video. Thanks for

378
00:16:35,040 --> 00:16:40,730
watching. Bye.

379
00:16:37,490 --> 00:16:40,730
[Music]

380
00:16:45,199 --> 00:16:50,959
Yeah. How awesome was that?

381
00:16:48,480 --> 00:16:54,560
Okay, so um unfortunately Arabella is

382
00:16:50,959 --> 00:16:57,199
notably not here, but I know you all

383
00:16:54,560 --> 00:17:00,480
have questions. And I have I have

384
00:16:57,199 --> 00:17:03,440
questions. Um you can post those

385
00:17:00,480 --> 00:17:05,760
questions in the Discord.

386
00:17:03,440 --> 00:17:08,799
She's not there yet, but she will be

387
00:17:05,760 --> 00:17:10,720
hopefully soon. Also, that's a we're

388
00:17:08,799 --> 00:17:12,079
going to have uh we've got one more

389
00:17:10,720 --> 00:17:14,559
video presentation in a second and then

390
00:17:12,079 --> 00:17:16,880
we're going to have live physical

391
00:17:14,559 --> 00:17:19,520
located students and then hold all of

392
00:17:16,880 --> 00:17:21,280
your questions until the end. I don't

393
00:17:19,520 --> 00:17:23,919
know about you, but I have really bad

394
00:17:21,280 --> 00:17:27,280
memory. So if you want you should write

395
00:17:23,919 --> 00:17:30,160
your questions down and then at the end

396
00:17:27,280 --> 00:17:32,559
we can ask them live put them in the

397
00:17:30,160 --> 00:17:34,240
discord so that yeah we know where they

398
00:17:32,559 --> 00:17:35,200
are and then the students can feel

399
00:17:34,240 --> 00:17:38,880
special because then they'll get

400
00:17:35,200 --> 00:17:42,240
questions. Um cool. Next up we have not

401
00:17:38,880 --> 00:17:43,760
Ally again we have Hudson. Um he is

402
00:17:42,240 --> 00:17:45,919
going to be talking about the fact that

403
00:17:43,760 --> 00:17:49,520
words aren't really no words are

404
00:17:45,919 --> 00:17:52,000
vectors. Um he's been programming he

405
00:17:49,520 --> 00:17:54,000
started programming in 2023 at the start

406
00:17:52,000 --> 00:17:57,760
of year 9 and it's been a bit over two

407
00:17:54,000 --> 00:18:02,640
years now. Um Hudson's favorite emoji is

408
00:17:57,760 --> 00:18:04,640
the chicken emoji. Yeah. Uh and he

409
00:18:02,640 --> 00:18:06,240
recently learned that in his computer

410
00:18:04,640 --> 00:18:08,240
science class we were all studying

411
00:18:06,240 --> 00:18:10,480
different storage techniques. I learned

412
00:18:08,240 --> 00:18:14,320
that theoretically all of the data that

413
00:18:10,480 --> 00:18:17,120
humans will generate in 2025 over 33

414
00:18:14,320 --> 00:18:20,080
zetabytes could be compressed with DNA

415
00:18:17,120 --> 00:18:21,760
storage into the size of a pingpong

416
00:18:20,080 --> 00:18:24,960
ball.

417
00:18:21,760 --> 00:18:27,280
And now we all know that it's wild.

418
00:18:24,960 --> 00:18:29,760
Okay. Uh Hudson.

419
00:18:27,280 --> 00:18:31,200
Hi Pon. I'm Hudson Wolfford. Let me ask

420
00:18:29,760 --> 00:18:33,840
you something. Have you ever wondered

421
00:18:31,200 --> 00:18:35,840
how chatbt can understand your prompts?

422
00:18:33,840 --> 00:18:37,919
Or maybe how Outlook categorizes your

423
00:18:35,840 --> 00:18:39,760
emails? Or maybe you've questioned how

424
00:18:37,919 --> 00:18:43,600
Google can tell the difference between

425
00:18:39,760 --> 00:18:46,000
Apple and an apple. Well, there is one

426
00:18:43,600 --> 00:18:48,240
thing in common with all these word

427
00:18:46,000 --> 00:18:49,520
embeds. In this video, I'll walk you

428
00:18:48,240 --> 00:18:51,440
through my journey learning about word

429
00:18:49,520 --> 00:18:53,840
embeds and how I use them in my game

430
00:18:51,440 --> 00:18:55,919
linkm, an online puzzle about chaining

431
00:18:53,840 --> 00:18:58,640
semantically similar words to reach a

432
00:18:55,919 --> 00:19:01,039
target word. But first, some background

433
00:18:58,640 --> 00:19:03,120
on me and the project. I started coding

434
00:19:01,039 --> 00:19:05,360
a little over two years ago when I first

435
00:19:03,120 --> 00:19:07,919
moved to Australia. This was mostly in

436
00:19:05,360 --> 00:19:10,240
unity with C where I made fun games with

437
00:19:07,919 --> 00:19:12,880
my classmates. But this year I started

438
00:19:10,240 --> 00:19:15,120
IB computer science and as part of the

439
00:19:12,880 --> 00:19:17,760
class we were given a project to create

440
00:19:15,120 --> 00:19:19,679
anything using Python. Naturally during

441
00:19:17,760 --> 00:19:21,440
this ideation process I got a little

442
00:19:19,679 --> 00:19:23,679
distracted playing word games like

443
00:19:21,440 --> 00:19:25,520
semantle or contextto. If you've never

444
00:19:23,679 --> 00:19:28,000
played these before, the goal is to

445
00:19:25,520 --> 00:19:30,000
guess a target word where each guess is

446
00:19:28,000 --> 00:19:31,440
given a similarity ranking to the

447
00:19:30,000 --> 00:19:33,039
targets.

448
00:19:31,440 --> 00:19:34,400
While playing these, it struck me how

449
00:19:33,039 --> 00:19:36,559
well the computer could understand

450
00:19:34,400 --> 00:19:38,640
language. And I immediately knew I

451
00:19:36,559 --> 00:19:40,480
wanted to do something similar. So, I

452
00:19:38,640 --> 00:19:43,520
researched how they work and figured out

453
00:19:40,480 --> 00:19:45,679
that they rely on word embeddings. Put

454
00:19:43,520 --> 00:19:48,480
simply, these embeddings are vector

455
00:19:45,679 --> 00:19:51,039
representations of words. For example,

456
00:19:48,480 --> 00:19:56,160
the word king may have the coordinates 2

457
00:19:51,039 --> 00:19:59,679
and 4. Queen may be at 2.5 and 3.5. And

458
00:19:56,160 --> 00:20:02,160
Python may be at four and one. These are

459
00:19:59,679 --> 00:20:04,799
meant to capture the meaning of words

460
00:20:02,160 --> 00:20:06,880
with words of similar meaning possessing

461
00:20:04,799 --> 00:20:08,880
similar vectors.

462
00:20:06,880 --> 00:20:10,559
Many algorithms exist to create these

463
00:20:08,880 --> 00:20:12,799
word embeddings, but the one that I

464
00:20:10,559 --> 00:20:15,039
found the most success with was global

465
00:20:12,799 --> 00:20:18,000
vectors for word representation or

466
00:20:15,039 --> 00:20:21,120
glove. What glove does is take some

467
00:20:18,000 --> 00:20:24,080
text, a corpus. Here's an example corpus

468
00:20:21,120 --> 00:20:26,799
from Charles Dickens. Next, Glove scans

469
00:20:24,080 --> 00:20:29,039
this to generate a co-occurrence matrix

470
00:20:26,799 --> 00:20:31,200
between each word. This shows the

471
00:20:29,039 --> 00:20:34,159
frequency that each word appears in

472
00:20:31,200 --> 00:20:36,480
another word's context. Then through

473
00:20:34,159 --> 00:20:38,960
matrix factorization, they generate

474
00:20:36,480 --> 00:20:41,520
vector representations of each word that

475
00:20:38,960 --> 00:20:43,360
demonstrate semantic meaning. To really

476
00:20:41,520 --> 00:20:46,320
capture the meaning of words, you need

477
00:20:43,360 --> 00:20:48,159
an enormous corpus. Thankfully, the

478
00:20:46,320 --> 00:20:50,320
Stanford Natural Language Processing

479
00:20:48,159 --> 00:20:52,159
Group offers pre-trained vectors on

480
00:20:50,320 --> 00:20:54,400
Wikipedia, so I didn't have to train

481
00:20:52,159 --> 00:20:56,320
these vectors myself. Now that I had

482
00:20:54,400 --> 00:20:58,640
these vectors, I could measure the

483
00:20:56,320 --> 00:21:01,120
distance between two words to tell how

484
00:20:58,640 --> 00:21:02,480
close they are in meaning. However, this

485
00:21:01,120 --> 00:21:05,039
isn't done with your conventional

486
00:21:02,480 --> 00:21:07,440
uklidian distance operation. Using this

487
00:21:05,039 --> 00:21:09,120
actually provides inaccurate results as

488
00:21:07,440 --> 00:21:11,280
meaning is more associated with the

489
00:21:09,120 --> 00:21:13,760
direction of the word from the origin

490
00:21:11,280 --> 00:21:16,320
rather than position. Instead, we take

491
00:21:13,760 --> 00:21:18,320
what's called a cosine similarity.

492
00:21:16,320 --> 00:21:20,559
Cosine similarity leverages the vector

493
00:21:18,320 --> 00:21:23,280
form of the words to measure the angle

494
00:21:20,559 --> 00:21:26,480
between them. Here let's call the angle

495
00:21:23,280 --> 00:21:29,039
between queen and king theta. Then we

496
00:21:26,480 --> 00:21:31,200
can take the coine of this angle.

497
00:21:29,039 --> 00:21:33,600
Varying the position of the word vectors

498
00:21:31,200 --> 00:21:35,919
varies the cosine as well with farther

499
00:21:33,600 --> 00:21:39,120
words providing lower values and closer

500
00:21:35,919 --> 00:21:40,880
words producing higher values. Of course

501
00:21:39,120 --> 00:21:43,280
cosine provides a value between negative

502
00:21:40,880 --> 00:21:46,080
1 and one. But some simple operations

503
00:21:43,280 --> 00:21:47,600
can remap this to between zero and one.

504
00:21:46,080 --> 00:21:50,720
And hence we now have a percentage

505
00:21:47,600 --> 00:21:53,520
similarity of different words. This is

506
00:21:50,720 --> 00:21:55,520
how we determine word similarity. I've

507
00:21:53,520 --> 00:21:57,440
simplified these vectors to 2D, but in

508
00:21:55,520 --> 00:21:59,520
reality, these vectors are often over 50

509
00:21:57,440 --> 00:22:02,159
dimensions. In fact, the ones that I use

510
00:21:59,520 --> 00:22:04,159
in linking them are 100 dimensional.

511
00:22:02,159 --> 00:22:06,640
Finding the angle between two of these

512
00:22:04,159 --> 00:22:08,799
seems complex, but this is where the dot

513
00:22:06,640 --> 00:22:11,520
product comes in. Let's take the dot

514
00:22:08,799 --> 00:22:14,640
product of two two-dimensional vectors A

515
00:22:11,520 --> 00:22:17,360
and B. Expanded. This is simply a sum of

516
00:22:14,640 --> 00:22:19,520
the products of each dimension. And this

517
00:22:17,360 --> 00:22:22,080
value is actually the coine of the angle

518
00:22:19,520 --> 00:22:24,559
between the two vectors multiplied by

519
00:22:22,080 --> 00:22:26,799
their magnitudes. And therefore the

520
00:22:24,559 --> 00:22:29,280
cosine similarity is simply the

521
00:22:26,799 --> 00:22:32,960
dotproduct over the product of the

522
00:22:29,280 --> 00:22:35,120
magnitude of the two vectors.

523
00:22:32,960 --> 00:22:37,520
Implementing this in code is extremely

524
00:22:35,120 --> 00:22:40,799
simple. Thanks to PyTorch, the function

525
00:22:37,520 --> 00:22:42,880
that I use is only three lines. Now that

526
00:22:40,799 --> 00:22:45,679
I had a system to compare words, I

527
00:22:42,880 --> 00:22:47,520
planned the core idea of my game. The

528
00:22:45,679 --> 00:22:49,200
user would be given a random start and

529
00:22:47,520 --> 00:22:51,200
target word and then they would have to

530
00:22:49,200 --> 00:22:52,880
chain their way to the target by using

531
00:22:51,200 --> 00:22:55,440
words sufficiently similar to their

532
00:22:52,880 --> 00:22:58,159
previous guess. For example, to get from

533
00:22:55,440 --> 00:23:02,240
protest to jaguar, a chain might be

534
00:22:58,159 --> 00:23:06,480
protest to fight to hunt to animal to

535
00:23:02,240 --> 00:23:08,159
cat to panther and then to jaguar.

536
00:23:06,480 --> 00:23:10,640
It wasn't long before I had a console

537
00:23:08,159 --> 00:23:12,640
based version of the game working. All I

538
00:23:10,640 --> 00:23:14,400
had to do was generate a start and

539
00:23:12,640 --> 00:23:16,640
target word and have the user

540
00:23:14,400 --> 00:23:19,039
continuously guess within a specified

541
00:23:16,640 --> 00:23:20,880
boundary of similarity.

542
00:23:19,039 --> 00:23:23,039
And there we go. That's how I made

543
00:23:20,880 --> 00:23:26,240
Lincoln. Thanks for watching and I

544
00:23:23,039 --> 00:23:30,440
appreciate your time.

545
00:23:26,240 --> 00:23:30,440
Just kidding. I still do. Great.

546
00:23:37,440 --> 00:23:40,679
I'm confused.

547
00:23:42,559 --> 00:23:49,159
a fake ending.

548
00:23:45,280 --> 00:23:49,159
Hang on a second.

549
00:23:50,400 --> 00:23:54,960
Create an entire UI for the project,

550
00:23:52,720 --> 00:23:55,679
which of course ended up taking the most

551
00:23:54,960 --> 00:23:57,200
time.

552
00:23:55,679 --> 00:23:59,520
Since we're only allowed to code in

553
00:23:57,200 --> 00:24:02,000
Python, I chose to use the nice GUI

554
00:23:59,520 --> 00:24:04,080
Python library. It uses your browser to

555
00:24:02,000 --> 00:24:06,400
render the UI, allowing me to test it in

556
00:24:04,080 --> 00:24:08,559
a website environment. A lot of this

557
00:24:06,400 --> 00:24:10,559
code is boring text and formatting, so I

558
00:24:08,559 --> 00:24:12,720
won't waste time going over it. However,

559
00:24:10,559 --> 00:24:14,559
there were some interesting parts. For

560
00:24:12,720 --> 00:24:16,480
example, I wanted to create a hidden

561
00:24:14,559 --> 00:24:18,320
system for the user where the computer

562
00:24:16,480 --> 00:24:19,600
would automatically solve the puzzle.

563
00:24:18,320 --> 00:24:21,279
While this initially seems

564
00:24:19,600 --> 00:24:23,840
straightforward, it turned out to be a

565
00:24:21,279 --> 00:24:26,159
complex path finding problem. Imagine

566
00:24:23,840 --> 00:24:27,919
each word as a node and a graph with

567
00:24:26,159 --> 00:24:30,159
edges representing possible connections

568
00:24:27,919 --> 00:24:32,159
in the game. In order to get from one

569
00:24:30,159 --> 00:24:33,840
word to another in the shortest time

570
00:24:32,159 --> 00:24:36,400
possible, you need to search the graph

571
00:24:33,840 --> 00:24:38,320
for an optimal path. Many algorithms

572
00:24:36,400 --> 00:24:40,640
exist to achieve this and I experimented

573
00:24:38,320 --> 00:24:42,880
with breathford search or dystras, but

574
00:24:40,640 --> 00:24:45,279
these were too slow. Instead, I wanted

575
00:24:42,880 --> 00:24:47,200
almost instant feedback for the user.

576
00:24:45,279 --> 00:24:48,720
So, I opted for a greedy algorithm.

577
00:24:47,200 --> 00:24:51,039
While it doesn't always provide the

578
00:24:48,720 --> 00:24:53,679
shortest path, that was my goal. I only

579
00:24:51,039 --> 00:24:56,000
wanted to find a path quickly. How the

580
00:24:53,679 --> 00:24:57,840
greedy algorithm works is to check all

581
00:24:56,000 --> 00:24:59,679
connections from the start node and

582
00:24:57,840 --> 00:25:03,360
calculate the similarity to the target

583
00:24:59,679 --> 00:25:05,039
for each one. This is the heristic. Then

584
00:25:03,360 --> 00:25:07,679
it chooses the node with the most

585
00:25:05,039 --> 00:25:09,919
similarity to the target and repeats the

586
00:25:07,679 --> 00:25:12,159
process. This algorithm finds a path

587
00:25:09,919 --> 00:25:14,000
almost immediately to the target word.

588
00:25:12,159 --> 00:25:16,320
And so at this point I had both a

589
00:25:14,000 --> 00:25:18,400
functional game for the user and a hint

590
00:25:16,320 --> 00:25:19,919
system if they got stuck. The only next

591
00:25:18,400 --> 00:25:22,080
step was to release the game to the

592
00:25:19,919 --> 00:25:24,640
public. To do this, I transferred the

593
00:25:22,080 --> 00:25:26,559
linkinm code onto my Raspberry Pi and

594
00:25:24,640 --> 00:25:28,960
connected it to a Cloudflare tunnel to

595
00:25:26,559 --> 00:25:30,799
host it on my own custom domain. And

596
00:25:28,960 --> 00:25:33,279
that's it. You can play linkin by

597
00:25:30,799 --> 00:25:35,279
visiting linkinm.com or by searching for

598
00:25:33,279 --> 00:25:37,520
it on Google. But in case you don't have

599
00:25:35,279 --> 00:25:39,279
the time, here's a short demonstration.

600
00:25:37,520 --> 00:25:41,200
Here is an example of a randomly

601
00:25:39,279 --> 00:25:43,600
generated linkin game where the goal is

602
00:25:41,200 --> 00:25:45,360
to get from respondent to eradicate in

603
00:25:43,600 --> 00:25:47,120
the shortest amount of links possible.

604
00:25:45,360 --> 00:25:49,200
The progress bar here represents how

605
00:25:47,120 --> 00:25:52,320
similar you are to eradicate. So we can

606
00:25:49,200 --> 00:25:54,240
see that respondent is only 14% similar.

607
00:25:52,320 --> 00:25:56,559
To start, you would guess a word that is

608
00:25:54,240 --> 00:25:58,960
similar to respondent but is also on the

609
00:25:56,559 --> 00:26:00,960
way to eradicate. For example, you might

610
00:25:58,960 --> 00:26:02,559
guess defendant.

611
00:26:00,960 --> 00:26:04,320
And you can see here that that links

612
00:26:02,559 --> 00:26:06,480
successfully. Of course, if I guess

613
00:26:04,320 --> 00:26:08,480
something like run, this has no

614
00:26:06,480 --> 00:26:11,760
association with defendant. So

615
00:26:08,480 --> 00:26:13,919
therefore, it doesn't link. When I play

616
00:26:11,760 --> 00:26:15,760
these, I like to think backwards. So

617
00:26:13,919 --> 00:26:18,080
when I think of eradicates, I think of

618
00:26:15,760 --> 00:26:21,120
virus. When I think of virus, I think of

619
00:26:18,080 --> 00:26:25,039
immunity. Immunity ties into defense. So

620
00:26:21,120 --> 00:26:26,960
from defendant, I would go defense.

621
00:26:25,039 --> 00:26:30,880
Immunity.

622
00:26:26,960 --> 00:26:33,919
From here, I could go vaccines.

623
00:26:30,880 --> 00:26:37,440
And then we're very close now. 65%. So I

624
00:26:33,919 --> 00:26:39,919
think with virus eradicates.

625
00:26:37,440 --> 00:26:41,679
And there we go. That's licking in.

626
00:26:39,919 --> 00:26:43,600
Linkin was a great project to learn

627
00:26:41,679 --> 00:26:45,919
about word embeddings and Python in

628
00:26:43,600 --> 00:26:48,080
general. In fact, I plan to continue it

629
00:26:45,919 --> 00:26:49,600
into the future. Specifically, I'd like

630
00:26:48,080 --> 00:26:51,840
to learn how to train my own word

631
00:26:49,600 --> 00:26:54,960
vectors and also improve the hint system

632
00:26:51,840 --> 00:26:56,799
to use less obscure words. Anyways, I

633
00:26:54,960 --> 00:26:59,170
hope you enjoyed this video and thanks

634
00:26:56,799 --> 00:27:07,650
for watching.

635
00:26:59,170 --> 00:27:07,650
[Applause]

636
00:27:07,919 --> 00:27:13,840
For reals this time

637
00:27:11,120 --> 00:27:19,279
that was that was really awesome. Um,

638
00:27:13,840 --> 00:27:20,960
Kayn up on deck. Kalin is next.

639
00:27:19,279 --> 00:27:25,440
And

640
00:27:20,960 --> 00:27:28,640
Kayn has been programming for two years.

641
00:27:25,440 --> 00:27:31,840
His favorite emoji

642
00:27:28,640 --> 00:27:33,679
is a sad face.

643
00:27:31,840 --> 00:27:34,480
Is it Is it sad or is it

644
00:27:33,679 --> 00:27:37,440
It's one of those.

645
00:27:34,480 --> 00:27:38,720
H I mean Oh, it's Yeah, it's kind of

646
00:27:37,440 --> 00:27:39,039
shock sad.

647
00:27:38,720 --> 00:27:41,039
Sad.

648
00:27:39,039 --> 00:27:44,960
Shock sad.

649
00:27:41,039 --> 00:27:47,679
Yeah. Um, Kayn recently learned that

650
00:27:44,960 --> 00:27:50,480
Facebook's like button was originally

651
00:27:47,679 --> 00:27:52,159
called an awesome button, but they

652
00:27:50,480 --> 00:27:55,799
changed it before launch as they thought

653
00:27:52,159 --> 00:27:55,799
it was too childish.

654
00:27:56,080 --> 00:28:01,600
Boo. Um, Kayn is going to be talking

655
00:28:00,080 --> 00:28:05,760
about

656
00:28:01,600 --> 00:28:09,039
designing for dual audiences a study app

657
00:28:05,760 --> 00:28:11,600
for students and teachers. Are you

658
00:28:09,039 --> 00:28:13,440
trying to teach your teachers something?

659
00:28:11,600 --> 00:28:18,200
Ah.

660
00:28:13,440 --> 00:28:18,200
Uh, Kayen, off you go. Thank you.

661
00:28:18,880 --> 00:28:23,960
Let's get a round of applause while I

662
00:28:20,559 --> 00:28:23,960
sort this out.

663
00:28:31,120 --> 00:28:36,080
Thank you. Of course. Of course. Um,

664
00:28:34,000 --> 00:28:40,799
yeah. This is Ajura, my study app for

665
00:28:36,080 --> 00:28:43,360
teachers and students combined. Um,

666
00:28:40,799 --> 00:28:44,799
slides aren't. There we go. Yeah. So,

667
00:28:43,360 --> 00:28:46,559
why did I build Ajura? Well, I built

668
00:28:44,799 --> 00:28:48,240
Ajura in the first place obviously to

669
00:28:46,559 --> 00:28:50,320
keep track of studying, homework, and

670
00:28:48,240 --> 00:28:52,000
assignments, but most importantly to

671
00:28:50,320 --> 00:28:53,520
link students and teachers because

672
00:28:52,000 --> 00:28:55,360
typically, you know, study apps like

673
00:28:53,520 --> 00:28:57,279
Canvas obviously are very teacher

674
00:28:55,360 --> 00:28:58,720
one-sided and so you you kind of have

675
00:28:57,279 --> 00:29:01,279
this disconnect between the teacher and

676
00:28:58,720 --> 00:29:02,960
the student. Um and you know obviously

677
00:29:01,279 --> 00:29:04,960
through canvas teachers only see

678
00:29:02,960 --> 00:29:06,720
underlying marks and trends through

679
00:29:04,960 --> 00:29:08,240
something like marks which are you know

680
00:29:06,720 --> 00:29:10,640
very one-dimensional and very hard to

681
00:29:08,240 --> 00:29:12,159
see. So this is where Ajura comes in and

682
00:29:10,640 --> 00:29:13,600
and that's one of my assignments that I

683
00:29:12,159 --> 00:29:16,720
submitted late because I didn't keep

684
00:29:13,600 --> 00:29:18,640
track of it. Um but yeah this is Ajura.

685
00:29:16,720 --> 00:29:20,880
So, Ajura is a platform designed for

686
00:29:18,640 --> 00:29:22,240
both students and teachers. Um, where

687
00:29:20,880 --> 00:29:24,240
educators have the ability to both

688
00:29:22,240 --> 00:29:25,840
create and manage classes, use a join

689
00:29:24,240 --> 00:29:27,600
code to individually invite students as

690
00:29:25,840 --> 00:29:30,320
well as assign tasks, monitor task

691
00:29:27,600 --> 00:29:31,760
completion, and see studying trends. But

692
00:29:30,320 --> 00:29:33,360
also for students, as you can see,

693
00:29:31,760 --> 00:29:34,880
that's a student dashboard up there.

694
00:29:33,360 --> 00:29:36,559
Students can log their study time

695
00:29:34,880 --> 00:29:38,399
visualized in a variety of graphs, as

696
00:29:36,559 --> 00:29:40,640
well as create and receive their own

697
00:29:38,399 --> 00:29:42,399
tasks. Um, and graphs for me just kind

698
00:29:40,640 --> 00:29:45,200
of help visualize my data because you

699
00:29:42,399 --> 00:29:46,559
often spend I think it's a total of, you

700
00:29:45,200 --> 00:29:47,600
know, close to a thousand hours over

701
00:29:46,559 --> 00:29:49,039
your school career or something like

702
00:29:47,600 --> 00:29:50,559
that studying. And so seeing that data

703
00:29:49,039 --> 00:29:52,640
and actually visualizing that data for

704
00:29:50,559 --> 00:29:54,159
me was really nice to see and kind of

705
00:29:52,640 --> 00:29:55,679
rewarding for all the studying I was

706
00:29:54,159 --> 00:29:58,240
doing, especially in the leadup to the

707
00:29:55,679 --> 00:30:00,320
HSC. Um, but yeah, so this is what I

708
00:29:58,240 --> 00:30:02,399
did. So my main issue was designing a UI

709
00:30:00,320 --> 00:30:04,720
for two different groups of users. And

710
00:30:02,399 --> 00:30:06,960
doing this, I used Flask and Ginger, um,

711
00:30:04,720 --> 00:30:09,679
which is a Python backend for serving

712
00:30:06,960 --> 00:30:11,600
routes. Um and then Ginger to

713
00:30:09,679 --> 00:30:12,559
dynamically render the HTML pages. As

714
00:30:11,600 --> 00:30:14,640
you can see there, that's just some

715
00:30:12,559 --> 00:30:16,799
simple code on how it works. Um but

716
00:30:14,640 --> 00:30:18,799
let's actually see what it looks like.

717
00:30:16,799 --> 00:30:21,120
Um so in creating the studentf facing

718
00:30:18,799 --> 00:30:23,039
interface, I am for an intuitive easy to

719
00:30:21,120 --> 00:30:24,720
easy to use UI which provided broad data

720
00:30:23,039 --> 00:30:27,039
visualizations for all their studying

721
00:30:24,720 --> 00:30:28,960
and tasks in one place. As you can see,

722
00:30:27,039 --> 00:30:30,960
this also involves joining and joining

723
00:30:28,960 --> 00:30:33,919
classes as well as allocating studying

724
00:30:30,960 --> 00:30:35,679
towards classes as you can see here. Um

725
00:30:33,919 --> 00:30:37,919
and then also creating study sessions

726
00:30:35,679 --> 00:30:39,919
towards classes which you can track your

727
00:30:37,919 --> 00:30:41,679
per day studying and viewing and

728
00:30:39,919 --> 00:30:45,279
completing tasks assigned by teachers

729
00:30:41,679 --> 00:30:47,840
and also created by yourself.

730
00:30:45,279 --> 00:30:50,720
Yeah. And you can sort them as well but

731
00:30:47,840 --> 00:30:52,320
move on. Uh yeah in comparison my

732
00:30:50,720 --> 00:30:54,320
teacher facing interface was designed to

733
00:30:52,320 --> 00:30:56,159
be more functional and data inensive

734
00:30:54,320 --> 00:30:58,480
supporting class creation and editing as

735
00:30:56,159 --> 00:31:00,720
well as detailed class views. Each class

736
00:30:58,480 --> 00:31:02,559
view included task, session, and student

737
00:31:00,720 --> 00:31:04,640
filtering along with interactive data

738
00:31:02,559 --> 00:31:06,720
visualizations to highlight trends and

739
00:31:04,640 --> 00:31:09,679
provide insights into what students were

740
00:31:06,720 --> 00:31:11,039
actually doing in real time. Also, you

741
00:31:09,679 --> 00:31:13,200
can invite students and kind of like

742
00:31:11,039 --> 00:31:15,360
Cahoot, you can join classes via join

743
00:31:13,200 --> 00:31:17,279
codes. Um, and this is just the teacher

744
00:31:15,360 --> 00:31:20,679
interface for managing tasks across

745
00:31:17,279 --> 00:31:20,679
their different classes.

746
00:31:20,720 --> 00:31:24,640
Video is going to keep going, but I'll

747
00:31:22,559 --> 00:31:26,240
move on. Um, you know, there were a

748
00:31:24,640 --> 00:31:28,399
couple of issues when I was creating my

749
00:31:26,240 --> 00:31:30,799
app. Primarily it was the tasking

750
00:31:28,399 --> 00:31:32,320
system. Um and while it doesn't seem

751
00:31:30,799 --> 00:31:33,760
like a big issue, I wanted students to

752
00:31:32,320 --> 00:31:35,840
have the ability to create their own

753
00:31:33,760 --> 00:31:39,919
tasks as well as for teachers to assign

754
00:31:35,840 --> 00:31:41,440
tasks sort of wide wide classwide. Um

755
00:31:39,919 --> 00:31:43,279
and so initially I thought this sort of

756
00:31:41,440 --> 00:31:45,039
like iterative single table-based

757
00:31:43,279 --> 00:31:46,320
approach would work in which a teacher

758
00:31:45,039 --> 00:31:48,159
would iterate through students in a

759
00:31:46,320 --> 00:31:50,399
class creating different tasks as you

760
00:31:48,159 --> 00:31:52,159
can see here. Um, however, teachers have

761
00:31:50,399 --> 00:31:54,320
no reliable way to track these tasks

762
00:31:52,159 --> 00:31:56,399
once they've been output. Um, since each

763
00:31:54,320 --> 00:31:58,240
one has sort of this unique ID and

764
00:31:56,399 --> 00:31:59,600
storing them in a list of unique IDs, it

765
00:31:58,240 --> 00:32:01,279
proves inefficient when you have to

766
00:31:59,600 --> 00:32:02,880
iterate through different tasks to edit

767
00:32:01,279 --> 00:32:05,519
them or delete them or update anything

768
00:32:02,880 --> 00:32:07,840
about the task. Um, and so instead I

769
00:32:05,519 --> 00:32:10,159
opted for a two tablebased solution um,

770
00:32:07,840 --> 00:32:13,039
in which teachers upload this sort of

771
00:32:10,159 --> 00:32:15,120
template task to a template task table

772
00:32:13,039 --> 00:32:16,720
and then students have the ability to

773
00:32:15,120 --> 00:32:19,039
connect with this task via their own

774
00:32:16,720 --> 00:32:20,720
task with a foreign key. So basically

775
00:32:19,039 --> 00:32:22,080
teachers will create a task and then

776
00:32:20,720 --> 00:32:24,320
this iterates through students in the

777
00:32:22,080 --> 00:32:26,320
class creating a placeholder task within

778
00:32:24,320 --> 00:32:28,240
the individual student table and then

779
00:32:26,320 --> 00:32:29,840
these individual tasks have no values on

780
00:32:28,240 --> 00:32:32,720
their own and so they they sort of get

781
00:32:29,840 --> 00:32:33,840
their data from the teacher task. Um and

782
00:32:32,720 --> 00:32:35,279
yeah this just sort of makes

783
00:32:33,840 --> 00:32:36,799
interpreting the data and displaying the

784
00:32:35,279 --> 00:32:38,799
data really easy because you can sort of

785
00:32:36,799 --> 00:32:39,919
join the two tables. Um and then

786
00:32:38,799 --> 00:32:41,760
basically all I have to do to display

787
00:32:39,919 --> 00:32:44,399
this data is iterate through a 2D list

788
00:32:41,760 --> 00:32:47,919
as you can see here. Um and yeah you can

789
00:32:44,399 --> 00:32:50,799
just see all the tasking data.

790
00:32:47,919 --> 00:32:52,960
Um and yeah, so yeah, as Steve Jobs once

791
00:32:50,799 --> 00:32:54,080
said, technology is nothing. What's more

792
00:32:52,960 --> 00:32:55,600
important is that you have faith in

793
00:32:54,080 --> 00:32:57,200
people and that if you give them tools,

794
00:32:55,600 --> 00:32:59,440
they'll do wonderful things with them.

795
00:32:57,200 --> 00:33:00,720
Um in creating Endura, I I aimed for

796
00:32:59,440 --> 00:33:02,480
teachers and students to have the

797
00:33:00,720 --> 00:33:05,279
ability to interface with one another

798
00:33:02,480 --> 00:33:08,240
and connect via tasking, students, uh

799
00:33:05,279 --> 00:33:10,080
student tracking, and classes. Um and

800
00:33:08,240 --> 00:33:11,200
then I guess I can do a demo if time

801
00:33:10,080 --> 00:33:13,840
allows for it.

802
00:33:11,200 --> 00:33:15,760
Yeah, cool. Um I'll just do

803
00:33:13,840 --> 00:33:18,480
You're like going home late, right?

804
00:33:15,760 --> 00:33:20,640
Yeah, do a demo.

805
00:33:18,480 --> 00:33:22,240
Um, I don't know if this will actually

806
00:33:20,640 --> 00:33:23,840
work cuz

807
00:33:22,240 --> 00:33:24,960
blame the technology.

808
00:33:23,840 --> 00:33:29,760
Yeah, I I don't know if it's going to

809
00:33:24,960 --> 00:33:33,240
work cuz I think it's um I don't know. I

810
00:33:29,760 --> 00:33:33,240
don't know if it's sharing.

811
00:33:36,480 --> 00:33:40,880
Um,

812
00:33:38,640 --> 00:33:43,039
maybe I'll just like end the

813
00:33:40,880 --> 00:33:44,480
Yes.

814
00:33:43,039 --> 00:33:47,039
URL.

815
00:33:44,480 --> 00:33:48,399
Oh, yeah. Um, you can go to it. I'm just

816
00:33:47,039 --> 00:33:51,519
on localhost right now, but you can

817
00:33:48,399 --> 00:33:54,000
actually just go to Not that one. Yeah,

818
00:33:51,519 --> 00:33:55,919
this one here. And this is sort of the

819
00:33:54,000 --> 00:33:56,880
landing page. Yeah. So, yeah, this is

820
00:33:55,919 --> 00:33:58,880
the landing page,

821
00:33:56,880 --> 00:34:00,640
kaleng.pythanyware.com.

822
00:33:58,880 --> 00:34:02,320
It should be responsive for mobile as

823
00:34:00,640 --> 00:34:03,760
well. Um, so you can view it on your

824
00:34:02,320 --> 00:34:06,000
phones if you really want to.

825
00:34:03,760 --> 00:34:06,960
Can someone put that on Discord?

826
00:34:06,000 --> 00:34:08,240
Yeah, thank you.

827
00:34:06,960 --> 00:34:09,919
Um, I'm just going to use the local host

828
00:34:08,240 --> 00:34:11,760
version because it's a bit quicker. Um,

829
00:34:09,919 --> 00:34:13,919
and also I have data pre-loaded onto the

830
00:34:11,760 --> 00:34:15,359
database there, but you can view it in

831
00:34:13,919 --> 00:34:17,359
real time if you want to on the

832
00:34:15,359 --> 00:34:18,800
kaleng.pythany.com.

833
00:34:17,359 --> 00:34:20,240
Um, but yeah, I'm logged into a teacher

834
00:34:18,800 --> 00:34:22,079
view right now. As you can see, you have

835
00:34:20,240 --> 00:34:24,240
a variety of classes. You can create

836
00:34:22,079 --> 00:34:26,639
your own class like this. Um, and you

837
00:34:24,240 --> 00:34:28,320
can also view your current class. And in

838
00:34:26,639 --> 00:34:30,399
here, you can create tasks, as you can

839
00:34:28,320 --> 00:34:32,480
see, and distribute these tasks to the

840
00:34:30,399 --> 00:34:34,800
class as well as sort of view all these

841
00:34:32,480 --> 00:34:36,159
people in the class. Um, I kind of Yeah,

842
00:34:34,800 --> 00:34:38,480
these are random names. Don't worry

843
00:34:36,159 --> 00:34:40,879
about them. Um you can order study time

844
00:34:38,480 --> 00:34:43,280
like so as well as view past sessions of

845
00:34:40,879 --> 00:34:44,879
what students have been studying. Um and

846
00:34:43,280 --> 00:34:47,359
then obviously just manage your tasks in

847
00:34:44,879 --> 00:34:49,040
a wider task view. Uh and then from a

848
00:34:47,359 --> 00:34:51,679
student view and this is kind of where

849
00:34:49,040 --> 00:34:53,040
the main sort of issue I faced was

850
00:34:51,679 --> 00:34:54,800
designing these two different views. The

851
00:34:53,040 --> 00:34:56,560
student view you can obviously see your

852
00:34:54,800 --> 00:34:58,560
study for this month and track it day by

853
00:34:56,560 --> 00:35:01,040
day as well as join classes similar to

854
00:34:58,560 --> 00:35:02,720
who with join codes uh and create and

855
00:35:01,040 --> 00:35:05,280
manage your own tasks like this as you

856
00:35:02,720 --> 00:35:10,200
can see personal versus assigned tasks.

857
00:35:05,280 --> 00:35:10,200
But yeah, that's it. Thank you.

858
00:35:10,320 --> 00:35:19,160
Thank you so much. Evelyn, you're up.

859
00:35:14,240 --> 00:35:19,160
All right. Nice work.

860
00:35:23,839 --> 00:35:27,720
Give me another round of applause.

861
00:35:29,920 --> 00:35:37,440
Okay, up on deck now we have Evelyn.

862
00:35:35,040 --> 00:35:42,119
Evelyn has been programming for three

863
00:35:37,440 --> 00:35:42,119
years, specifically three years.

864
00:35:44,000 --> 00:35:48,240
I assume that's a loadbearing question

865
00:35:45,839 --> 00:35:52,240
mark. Um,

866
00:35:48,240 --> 00:35:55,760
uh, Evelyn recently learned about Temple

867
00:35:52,240 --> 00:35:57,839
OS, an operating system written

868
00:35:55,760 --> 00:36:00,800
Oh, yeah.

869
00:35:57,839 --> 00:36:03,680
entirely by one guy, Terry Davis, who

870
00:36:00,800 --> 00:36:07,520
also wrote the language it was coded in

871
00:36:03,680 --> 00:36:11,119
called Holy Sea. Uh, and Evelyn's

872
00:36:07,520 --> 00:36:14,400
favorite emoji is I really should have

873
00:36:11,119 --> 00:36:17,280
bumped the font of this particular

874
00:36:14,400 --> 00:36:21,280
uh Oh,

875
00:36:17,280 --> 00:36:23,520
it's the here no evil monkey. I like it.

876
00:36:21,280 --> 00:36:25,780
That's Evelyn's favorite emoji. Evelyn,

877
00:36:23,520 --> 00:36:30,929
thank you very much.

878
00:36:25,780 --> 00:36:30,929
[Applause]

879
00:36:31,040 --> 00:36:37,760
Um, hi. Is this working? Okay. Um, hi.

880
00:36:35,359 --> 00:36:40,960
I'm Evelyn and this is my presentation

881
00:36:37,760 --> 00:36:42,560
on using RMS and STFTs to visualize

882
00:36:40,960 --> 00:36:44,960
music.

883
00:36:42,560 --> 00:36:50,000
Um, to begin with, I'm a year 12 student

884
00:36:44,960 --> 00:36:52,160
at Bergman Anglican School. This is Oh,

885
00:36:50,000 --> 00:36:54,560
it's in the wrong order. A fun fact

886
00:36:52,160 --> 00:36:57,280
about me is that I play piano. And this

887
00:36:54,560 --> 00:36:59,680
is my self-portrait.

888
00:36:57,280 --> 00:37:02,079
Um, my project is a music visualizer.

889
00:36:59,680 --> 00:37:04,240
More specifically, it uses shorttime for

890
00:37:02,079 --> 00:37:07,440
transform to break down and analyze a

891
00:37:04,240 --> 00:37:09,440
song or any audio file really. Um, it

892
00:37:07,440 --> 00:37:11,119
plays this audio alongside a sine wave

893
00:37:09,440 --> 00:37:13,760
that changes amplitude based on the

894
00:37:11,119 --> 00:37:16,640
amplitudes of various frequencies at a

895
00:37:13,760 --> 00:37:18,720
given point in the song.

896
00:37:16,640 --> 00:37:20,240
So, why did I choose to make a music

897
00:37:18,720 --> 00:37:24,480
visualizer?

898
00:37:20,240 --> 00:37:26,079
Firstly, I like music a lot. Um, I'm

899
00:37:24,480 --> 00:37:27,839
really glad that I was able to come up

900
00:37:26,079 --> 00:37:29,920
with a project idea that combined my

901
00:37:27,839 --> 00:37:32,160
interest so well and it has made the

902
00:37:29,920 --> 00:37:35,760
whole process overall a lot more fun and

903
00:37:32,160 --> 00:37:39,760
rewarding. And secondly, I like pretty

904
00:37:35,760 --> 00:37:41,839
pictures. I mean, who doesn't? Um,

905
00:37:39,760 --> 00:37:43,520
before I get into the details, these are

906
00:37:41,839 --> 00:37:46,560
some important terms and their

907
00:37:43,520 --> 00:37:49,200
definitions. A time series is an audio

908
00:37:46,560 --> 00:37:51,119
signal denoted by Y, where Y of T

909
00:37:49,200 --> 00:37:55,280
corresponds to the amplitude of a

910
00:37:51,119 --> 00:37:57,680
waveform at sample t. A sample is a

911
00:37:55,280 --> 00:37:59,920
value of a signal at a point in time.

912
00:37:57,680 --> 00:38:02,320
The sample rate is the number of samples

913
00:37:59,920 --> 00:38:06,400
per second of a time series and defaults

914
00:38:02,320 --> 00:38:08,400
to 22,50 in my use case. And a frame is

915
00:38:06,400 --> 00:38:10,880
a short slice of a time series used for

916
00:38:08,400 --> 00:38:13,200
analysis purposes. The length of a frame

917
00:38:10,880 --> 00:38:18,000
in terms of samples is an integer

918
00:38:13,200 --> 00:38:20,960
variable and defaults to48 in my library

919
00:38:18,000 --> 00:38:23,680
that I used not that I made.

920
00:38:20,960 --> 00:38:26,079
Um just very quickly RMS is what I used

921
00:38:23,680 --> 00:38:28,079
for the sort of prototype of my project.

922
00:38:26,079 --> 00:38:30,000
It stands for root mean square and is

923
00:38:28,079 --> 00:38:32,079
calculated by squaring the amplitudes of

924
00:38:30,000 --> 00:38:33,760
a signal in a frame taking the average

925
00:38:32,079 --> 00:38:36,640
of that and then square rooting that

926
00:38:33,760 --> 00:38:38,880
average. This results in a sign a single

927
00:38:36,640 --> 00:38:41,200
sine wave that increases in amplitude if

928
00:38:38,880 --> 00:38:44,079
the music is louder and vice versa if

929
00:38:41,200 --> 00:38:46,160
the music is softer.

930
00:38:44,079 --> 00:38:48,079
Now onto the current version of my

931
00:38:46,160 --> 00:38:50,880
project that uses shorttime for

932
00:38:48,079 --> 00:38:52,800
transform. So short-time for transform

933
00:38:50,880 --> 00:38:55,119
is a technique that involves computing

934
00:38:52,800 --> 00:38:57,359
for transforms over a lot of small

935
00:38:55,119 --> 00:38:59,119
equally long windows of signal data.

936
00:38:57,359 --> 00:39:01,599
These transforms break the data down

937
00:38:59,119 --> 00:39:03,040
into its components. that is they tell

938
00:39:01,599 --> 00:39:04,720
you the frequencies and their

939
00:39:03,040 --> 00:39:07,040
corresponding amplitudes that make up

940
00:39:04,720 --> 00:39:09,920
the data. In the context of the library

941
00:39:07,040 --> 00:39:12,880
that I used called Labrosa, it returns a

942
00:39:09,920 --> 00:39:14,640
complex valued matrix denoted D where

943
00:39:12,880 --> 00:39:17,440
the absolute value of it is the

944
00:39:14,640 --> 00:39:20,079
magnitude of frequency bin F at frame T

945
00:39:17,440 --> 00:39:23,520
where frequency bin is just the segment

946
00:39:20,079 --> 00:39:25,839
of the frequency axis.

947
00:39:23,520 --> 00:39:27,599
This is the equation of a forer

948
00:39:25,839 --> 00:39:30,880
transform.

949
00:39:27,599 --> 00:39:32,880
um I don't really understand it but even

950
00:39:30,880 --> 00:39:34,960
though I don't understand how it works

951
00:39:32,880 --> 00:39:37,359
exactly I have been able to understand

952
00:39:34,960 --> 00:39:39,680
what they do and how I can input things

953
00:39:37,359 --> 00:39:42,160
in to get things I want out and then use

954
00:39:39,680 --> 00:39:46,079
that in my project.

955
00:39:42,160 --> 00:39:48,720
So this is how I use STF in my program.

956
00:39:46,079 --> 00:39:50,480
Um I get the complex valued matrix here

957
00:39:48,720 --> 00:39:53,119
and then absolute it so I get the

958
00:39:50,480 --> 00:39:54,720
magnitudes of different frequencies.

959
00:39:53,119 --> 00:39:56,480
Then using another function from the

960
00:39:54,720 --> 00:39:58,400
library, I convert it to a decibel

961
00:39:56,480 --> 00:40:01,520
matrix so the values are better to work

962
00:39:58,400 --> 00:40:03,440
with. Um then what I've done is split

963
00:40:01,520 --> 00:40:06,720
the data into five groups based on

964
00:40:03,440 --> 00:40:08,880
frequencies. Basically just made more

965
00:40:06,720 --> 00:40:11,119
frequency bins and put the frequency

966
00:40:08,880 --> 00:40:12,960
bins into the frequency bins. Um and

967
00:40:11,119 --> 00:40:14,640
then for each frame I get the average

968
00:40:12,960 --> 00:40:16,480
plus the minimum amplitude in the

969
00:40:14,640 --> 00:40:18,079
original matrix. So all the values are

970
00:40:16,480 --> 00:40:20,160
positive because amplitudes can be

971
00:40:18,079 --> 00:40:21,760
negative. Um, I did this so I wouldn't

972
00:40:20,160 --> 00:40:26,800
have to deal with every piece of data

973
00:40:21,760 --> 00:40:29,119
from all 1,025 frequency bins.

974
00:40:26,800 --> 00:40:31,040
Um, then using these sections, this

975
00:40:29,119 --> 00:40:32,880
function will draw a sine wave.

976
00:40:31,040 --> 00:40:35,040
Essentially, I put the sections into

977
00:40:32,880 --> 00:40:36,960
another list called all amplitudes. Then

978
00:40:35,040 --> 00:40:38,960
I get the relevant amplitudes of each

979
00:40:36,960 --> 00:40:41,520
frequency group based on the position of

980
00:40:38,960 --> 00:40:44,079
music you're at. Um, and put those in a

981
00:40:41,520 --> 00:40:45,839
list. I then feed that list into the

982
00:40:44,079 --> 00:40:48,400
sine wave drawing function on the right

983
00:40:45,839 --> 00:40:50,400
there. uh the program will run until

984
00:40:48,400 --> 00:40:54,000
completion of the song or if you just

985
00:40:50,400 --> 00:40:55,920
manually exit the Pygame window.

986
00:40:54,000 --> 00:40:58,000
Something interesting I learned or kind

987
00:40:55,920 --> 00:40:59,760
of just noticed through testing was that

988
00:40:58,000 --> 00:41:01,680
the lower frequencies were generally a

989
00:40:59,760 --> 00:41:04,160
lot more active, especially in purely

990
00:41:01,680 --> 00:41:06,319
instrumental parts of a song. Um to

991
00:41:04,160 --> 00:41:08,560
research this in more concrete terms, I

992
00:41:06,319 --> 00:41:12,560
turned to the piano. The lowest note on

993
00:41:08,560 --> 00:41:15,599
a piano A0 has a frequency of 27.5 htz

994
00:41:12,560 --> 00:41:18,599
and the highest C8 has a frequency of

995
00:41:15,599 --> 00:41:18,599
4,186.01

996
00:41:18,720 --> 00:41:24,240
hertz. Using the Nyquist Shannon theorem

997
00:41:22,319 --> 00:41:25,760
um we understand that the maximum

998
00:41:24,240 --> 00:41:28,400
frequency that can be accurately

999
00:41:25,760 --> 00:41:31,200
represented as a digital signal is half

1000
00:41:28,400 --> 00:41:33,200
of the sample rate of the signal. Given

1001
00:41:31,200 --> 00:41:36,079
that the sample rate in my case was

1002
00:41:33,200 --> 00:41:40,160
22,50, the range of frequencies that

1003
00:41:36,079 --> 00:41:42,319
data was collected for was 0 to 11,025.

1004
00:41:40,160 --> 00:41:44,560
So it all kind of makes sense. If the

1005
00:41:42,319 --> 00:41:46,880
piano keys and thus instrumental parts

1006
00:41:44,560 --> 00:41:48,960
of a song only reach around 4,000 hertz

1007
00:41:46,880 --> 00:41:51,359
at max and the entire line is

1008
00:41:48,960 --> 00:41:54,400
representative of frequencies from 0 to

1009
00:41:51,359 --> 00:41:58,240
11,000, instrumentals would just sit

1010
00:41:54,400 --> 00:42:00,400
around the bottom half of the line.

1011
00:41:58,240 --> 00:42:02,560
Um, then when vocals came in, the upper

1012
00:42:00,400 --> 00:42:04,640
end of the line became more active,

1013
00:42:02,560 --> 00:42:07,119
which also kind of makes sense because

1014
00:42:04,640 --> 00:42:10,079
our voices naturally contain those high

1015
00:42:07,119 --> 00:42:12,960
frequency pitches. Of course, this is

1016
00:42:10,079 --> 00:42:15,920
like grossly oversimplified um because

1017
00:42:12,960 --> 00:42:18,720
sound as a note on an instrument uh is

1018
00:42:15,920 --> 00:42:21,920
going to consist of various frequencies,

1019
00:42:18,720 --> 00:42:25,119
but the general idea is there.

1020
00:42:21,920 --> 00:42:27,359
Uh is there a correlation here? Yes, the

1021
00:42:25,119 --> 00:42:29,920
data and theory appear to match up and

1022
00:42:27,359 --> 00:42:32,640
there's not really any big issues. Is

1023
00:42:29,920 --> 00:42:36,079
this proof of causation though? I don't

1024
00:42:32,640 --> 00:42:39,200
know. Um I can't know for sure. This is

1025
00:42:36,079 --> 00:42:41,200
just um from research that I've put two

1026
00:42:39,200 --> 00:42:44,640
things together. But I think it was an

1027
00:42:41,200 --> 00:42:46,319
interesting investigation nonetheless.

1028
00:42:44,640 --> 00:42:47,680
Uh moving forward, there are still many

1029
00:42:46,319 --> 00:42:49,920
things I want to keep adding to this

1030
00:42:47,680 --> 00:42:52,079
project. Firstly, for assignment

1031
00:42:49,920 --> 00:42:54,640
purposes, my teacher's forcing me to

1032
00:42:52,079 --> 00:42:56,880
include a UI.

1033
00:42:54,640 --> 00:42:58,560
Why would I need that? The user can just

1034
00:42:56,880 --> 00:43:01,560
enter the file name directly into the

1035
00:42:58,560 --> 00:43:01,560
code.

1036
00:43:03,200 --> 00:43:08,240
And also, I'd like to figure out a

1037
00:43:05,200 --> 00:43:10,240
better way of visualizing um the music.

1038
00:43:08,240 --> 00:43:12,720
And with this, then I can represent all

1039
00:43:10,240 --> 00:43:17,680
or a lot more frequencies.

1040
00:43:12,720 --> 00:43:18,880
Um, I can give a live demo now

1041
00:43:17,680 --> 00:43:19,520
if

1042
00:43:18,880 --> 00:43:21,040
demo

1043
00:43:19,520 --> 00:43:24,040
it wants to work.

1044
00:43:21,040 --> 00:43:24,040
Yes.

1045
00:43:28,450 --> 00:43:42,240
[Music]

1046
00:43:39,119 --> 00:43:43,920
It's been raining. I swear I'm not

1047
00:43:42,240 --> 00:43:47,839
lonely.

1048
00:43:43,920 --> 00:43:49,920
Always complaining that you never see

1049
00:43:47,839 --> 00:43:54,790
me.

1050
00:43:49,920 --> 00:44:00,719
Um, yes, that's it. Thank you.

1051
00:43:54,790 --> 00:44:00,719
[Applause]

1052
00:44:03,680 --> 00:44:07,359
I was too busy taking a photo of like

1053
00:44:05,680 --> 00:44:09,520
demos that work because it makes me

1054
00:44:07,359 --> 00:44:11,680
excited every single time. Evelyn, come

1055
00:44:09,520 --> 00:44:14,880
and give me a Hang on. Yes. Microphone

1056
00:44:11,680 --> 00:44:17,920
here. High five here. Excellent. Give

1057
00:44:14,880 --> 00:44:21,079
Evelyn a massive round of applause.

1058
00:44:17,920 --> 00:44:21,079
Thank you.

1059
00:44:21,760 --> 00:44:24,800
Not like this.

1060
00:44:22,480 --> 00:44:30,160
Like that.

1061
00:44:24,800 --> 00:44:32,079
Okay, next up we have

1062
00:44:30,160 --> 00:44:38,960
Mars.

1063
00:44:32,079 --> 00:44:42,720
Uh here Mars is going to talk about

1064
00:44:38,960 --> 00:44:45,440
Bartholomew, the Roman numeral

1065
00:44:42,720 --> 00:44:47,040
13.

1066
00:44:45,440 --> 00:44:49,760
Had to think about that for a second.

1067
00:44:47,040 --> 00:44:53,760
Um, perhaps related, Mars's uh favorite

1068
00:44:49,760 --> 00:44:56,560
emoji is the troll emoji. Uh, Mars has

1069
00:44:53,760 --> 00:44:59,520
been programming for 5 years and Mars

1070
00:44:56,560 --> 00:45:03,359
learned about ray tracing and how it

1071
00:44:59,520 --> 00:45:06,400
actually works. Brackets just a bunch of

1072
00:45:03,359 --> 00:45:08,720
maths and intersections.

1073
00:45:06,400 --> 00:45:12,240
Computers. computers,

1074
00:45:08,720 --> 00:45:16,120
just a bunch of maths. I like it. U

1075
00:45:12,240 --> 00:45:19,440
Mars, uh applause.

1076
00:45:16,120 --> 00:45:21,839
[Applause]

1077
00:45:19,440 --> 00:45:23,599
Thank you guys. There is a lot more

1078
00:45:21,839 --> 00:45:26,880
people than I thought there were going

1079
00:45:23,599 --> 00:45:29,280
to be. Okay. Um hi everyone. My name is

1080
00:45:26,880 --> 00:45:31,040
Mars and this is Bartholomew the 13th,

1081
00:45:29,280 --> 00:45:34,720
my customuilt convolutional neural

1082
00:45:31,040 --> 00:45:37,440
network. Um most people train or yeah

1083
00:45:34,720 --> 00:45:39,839
train CNN's using TensorFlow or PyTorch.

1084
00:45:37,440 --> 00:45:41,680
I built mine entirely in raw Python and

1085
00:45:39,839 --> 00:45:44,960
Numpy.

1086
00:45:41,680 --> 00:45:47,760
Why? Because I hate myself apparently.

1087
00:45:44,960 --> 00:45:53,119
Um but first a little bit more about

1088
00:45:47,760 --> 00:45:54,880
myself. My my name is Mars Karola. I am

1089
00:45:53,119 --> 00:45:57,760
a student at Bergman Anglican School

1090
00:45:54,880 --> 00:46:00,000
ACT. Um I am going to AU for advanced

1091
00:45:57,760 --> 00:46:01,920
computing. There are three facts that

1092
00:46:00,000 --> 00:46:04,640
everyone needs to know about me.

1093
00:46:01,920 --> 00:46:08,400
I built my first I built my first Python

1094
00:46:04,640 --> 00:46:11,200
project and in Python obviously

1095
00:46:08,400 --> 00:46:14,880
and it was a calculator. Apparently 1 +

1096
00:46:11,200 --> 00:46:19,599
1 is equal to 11.

1097
00:46:14,880 --> 00:46:22,599
Um I love numpy and I hate tomatoes.

1098
00:46:19,599 --> 00:46:22,599
Yeah.

1099
00:46:22,720 --> 00:46:26,400
All right. So what is Bartholomew?

1100
00:46:24,640 --> 00:46:28,240
Bartholomew as I said before is a

1101
00:46:26,400 --> 00:46:30,560
convolutional neural network built to

1102
00:46:28,240 --> 00:46:32,880
read handwriting. Now, something CNN's

1103
00:46:30,560 --> 00:46:35,680
can compare to are very picky art

1104
00:46:32,880 --> 00:46:38,000
critics.

1105
00:46:35,680 --> 00:46:40,319
It just doesn't look at an image. It

1106
00:46:38,000 --> 00:46:42,800
analyzes it for shapes, features, and

1107
00:46:40,319 --> 00:46:46,319
textures layer by layer until it can say

1108
00:46:42,800 --> 00:46:48,560
with confidence, yes, that is a seven.

1109
00:46:46,319 --> 00:46:50,640
So, the network learns to recognize

1110
00:46:48,560 --> 00:46:53,280
digit patterns by extracting in

1111
00:46:50,640 --> 00:46:54,960
important features from input images

1112
00:46:53,280 --> 00:46:57,119
through multiple layers, including

1113
00:46:54,960 --> 00:46:59,200
convolution layers, activation layers,

1114
00:46:57,119 --> 00:47:00,480
reshaping, and tense layers. It's

1115
00:46:59,200 --> 00:47:02,079
trained using forward and backwards

1116
00:47:00,480 --> 00:47:04,319
propagation algorithms which optimize

1117
00:47:02,079 --> 00:47:07,359
the network's parameters to improve

1118
00:47:04,319 --> 00:47:10,240
prediction. Oh

1119
00:47:07,359 --> 00:47:11,599
that improve prediction accuracy.

1120
00:47:10,240 --> 00:47:13,920
But first let's start with what exactly

1121
00:47:11,599 --> 00:47:15,920
a layer is in this broad context of

1122
00:47:13,920 --> 00:47:18,240
neural networks.

1123
00:47:15,920 --> 00:47:19,920
So a layer is essentially a big function

1124
00:47:18,240 --> 00:47:21,520
and therefore we can represent it as a

1125
00:47:19,920 --> 00:47:22,720
box.

1126
00:47:21,520 --> 00:47:24,079
Now something that's true for every

1127
00:47:22,720 --> 00:47:26,800
single layer is that if you give it an

1128
00:47:24,079 --> 00:47:29,440
input you will always get an output.

1129
00:47:26,800 --> 00:47:31,040
That is called forward propagation.

1130
00:47:29,440 --> 00:47:34,640
For each layer, there is always an input

1131
00:47:31,040 --> 00:47:37,280
and output. Easy enough, right? The

1132
00:47:34,640 --> 00:47:39,680
other step when training CNN's is called

1133
00:47:37,280 --> 00:47:41,440
backwards propagation. This is when that

1134
00:47:39,680 --> 00:47:43,200
layer updates its parameters. It's how

1135
00:47:41,440 --> 00:47:44,720
the network learns from its mistakes by

1136
00:47:43,200 --> 00:47:46,720
adjusting its weights and biases to

1137
00:47:44,720 --> 00:47:48,880
improve performance.

1138
00:47:46,720 --> 00:47:51,280
Inside these layers are the weights. The

1139
00:47:48,880 --> 00:47:54,079
weights decide how much influence each

1140
00:47:51,280 --> 00:47:55,920
pixel has. If the networks if the

1141
00:47:54,079 --> 00:47:57,680
network gets something wrong, we notch

1142
00:47:55,920 --> 00:47:59,920
the weight slightly. Thousands of

1143
00:47:57,680 --> 00:48:01,680
notches later, it gets a little bit

1144
00:47:59,920 --> 00:48:05,839
smarter.

1145
00:48:01,680 --> 00:48:09,359
Now, this is Bolu's entire um blueprint.

1146
00:48:05,839 --> 00:48:11,200
If it looks a bit um much right now,

1147
00:48:09,359 --> 00:48:13,200
don't worry. It's just four layers on

1148
00:48:11,200 --> 00:48:16,000
repeat. It's just convolutional layers,

1149
00:48:13,200 --> 00:48:18,480
activation layers, reshape layers, and

1150
00:48:16,000 --> 00:48:20,960
dense layers. So let's start with

1151
00:48:18,480 --> 00:48:23,520
convolutional layers since it's the

1152
00:48:20,960 --> 00:48:26,559
point of the entire project. Say we're

1153
00:48:23,520 --> 00:48:28,960
given two matrices, one that is 3x3 in

1154
00:48:26,559 --> 00:48:32,960
size and one is and one that is 2x2 in

1155
00:48:28,960 --> 00:48:35,119
size. The 2x2 one is called a kernel. In

1156
00:48:32,960 --> 00:48:36,559
the convolutional process, we slide the

1157
00:48:35,119 --> 00:48:38,880
kernel over the image performing

1158
00:48:36,559 --> 00:48:40,319
multiplication at each step to highlight

1159
00:48:38,880 --> 00:48:42,960
important patterns like edges or

1160
00:48:40,319 --> 00:48:45,359
textures. This process helps the network

1161
00:48:42,960 --> 00:48:47,680
find important details in the image like

1162
00:48:45,359 --> 00:48:49,760
edges or shapes. These details are

1163
00:48:47,680 --> 00:48:51,119
called features and they help the

1164
00:48:49,760 --> 00:48:52,960
network understand what's actually in

1165
00:48:51,119 --> 00:48:56,400
the image like recognizing handwritten

1166
00:48:52,960 --> 00:48:58,000
digits. In short, convolutions let the

1167
00:48:56,400 --> 00:48:59,839
network automatically pick out useful

1168
00:48:58,000 --> 00:49:01,839
patterns from raw images so it can

1169
00:48:59,839 --> 00:49:04,240
classify them correctly and learn from

1170
00:49:01,839 --> 00:49:06,079
these classifications and classify a

1171
00:49:04,240 --> 00:49:08,800
digit.

1172
00:49:06,079 --> 00:49:09,920
Next up is the activation layer. This is

1173
00:49:08,800 --> 00:49:12,319
where we add a little bit of

1174
00:49:09,920 --> 00:49:14,559
personality. Without it, the network is

1175
00:49:12,319 --> 00:49:16,880
just a boring linear model. The

1176
00:49:14,559 --> 00:49:18,640
activation layer applies a mathematical

1177
00:49:16,880 --> 00:49:21,040
function to the output of the previous

1178
00:49:18,640 --> 00:49:23,920
layer, introducing nonlinearity into the

1179
00:49:21,040 --> 00:49:25,760
network. Without activation functions,

1180
00:49:23,920 --> 00:49:29,839
the network would not be able to

1181
00:49:25,760 --> 00:49:32,319
classify any image. It just say it say

1182
00:49:29,839 --> 00:49:35,040
the same thing over and over again. Um,

1183
00:49:32,319 --> 00:49:37,599
I use sig mode for my one. Don't ask me

1184
00:49:35,040 --> 00:49:39,520
how that works because I have no clue at

1185
00:49:37,599 --> 00:49:42,000
all.

1186
00:49:39,520 --> 00:49:43,920
Um, and the final one is the dense

1187
00:49:42,000 --> 00:49:45,839
layer. The dense layer, also called a

1188
00:49:43,920 --> 00:49:47,520
fully connected layer, takes in all of

1189
00:49:45,839 --> 00:49:49,359
the features learned by earlier layers

1190
00:49:47,520 --> 00:49:51,680
and uses them to make a final decision

1191
00:49:49,359 --> 00:49:53,280
about what the image actually is. Every

1192
00:49:51,680 --> 00:49:55,599
neuron in this layer is connected to all

1193
00:49:53,280 --> 00:49:57,599
the neurons from the previous layer. And

1194
00:49:55,599 --> 00:50:00,240
it adds up all the inputs, weighs them,

1195
00:49:57,599 --> 00:50:01,760
and adds in a little bit of bias. Uh,

1196
00:50:00,240 --> 00:50:03,520
and then at the end, it uses an

1197
00:50:01,760 --> 00:50:06,400
activation function to actually classify

1198
00:50:03,520 --> 00:50:08,880
what it is. It's basically a decision

1199
00:50:06,400 --> 00:50:12,319
maker. Based on all these features, I'm

1200
00:50:08,880 --> 00:50:14,800
95% sure this is a three or sometimes

1201
00:50:12,319 --> 00:50:16,800
it's definitely a cat. What is a bit

1202
00:50:14,800 --> 00:50:19,520
awkward since my data set has no cats in

1203
00:50:16,800 --> 00:50:22,240
it.

1204
00:50:19,520 --> 00:50:24,960
So training for bathon mu is two steps

1205
00:50:22,240 --> 00:50:28,079
on repeat forward pass and backwards

1206
00:50:24,960 --> 00:50:30,160
pass. Uh it normally just predicts gets

1207
00:50:28,079 --> 00:50:33,520
it wrong and you fix the weights from

1208
00:50:30,160 --> 00:50:35,680
that predict. Repeat this for around 40

1209
00:50:33,520 --> 00:50:38,079
times over the entire data set. And for

1210
00:50:35,680 --> 00:50:40,319
10 hours, Bartholomew learns digits with

1211
00:50:38,079 --> 00:50:42,400
95% accuracy.

1212
00:50:40,319 --> 00:50:44,480
So after all that, Bartholomew can now

1213
00:50:42,400 --> 00:50:47,599
read handwriting almost as well as me on

1214
00:50:44,480 --> 00:50:49,440
3 hours of sleep.

1215
00:50:47,599 --> 00:50:51,359
So I made this model for one thing and

1216
00:50:49,440 --> 00:50:53,680
one thing only,

1217
00:50:51,359 --> 00:50:56,880
to read my handwriting. And I am proud

1218
00:50:53,680 --> 00:50:58,880
to say that after 10 hours of

1219
00:50:56,880 --> 00:51:01,200
mind-numbing training, it still can't

1220
00:50:58,880 --> 00:51:02,880
read my handwriting.

1221
00:51:01,200 --> 00:51:04,400
It might just be my handwriting. I don't

1222
00:51:02,880 --> 00:51:06,800
know.

1223
00:51:04,400 --> 00:51:09,440
So, I made a little GUI as you can see

1224
00:51:06,800 --> 00:51:12,000
on the board uh to show button view off.

1225
00:51:09,440 --> 00:51:16,800
You press a button and it grabs 20 um

1226
00:51:12,000 --> 00:51:19,119
random test images alongside its um true

1227
00:51:16,800 --> 00:51:22,079
labels and the predicted labels. Um

1228
00:51:19,119 --> 00:51:24,559
Bartholomew is quite accurate, but

1229
00:51:22,079 --> 00:51:27,119
sometimes it gets things wrong. It

1230
00:51:24,559 --> 00:51:31,440
thought it was a two and a nine instead

1231
00:51:27,119 --> 00:51:34,640
of sevens. It's pretty stupid.

1232
00:51:31,440 --> 00:51:36,960
So now for the fun part, the challenges.

1233
00:51:34,640 --> 00:51:39,040
First, I had overfitting. Both me was

1234
00:51:36,960 --> 00:51:41,200
acing the training set but flunking the

1235
00:51:39,040 --> 00:51:42,240
test data. I fixed it with shuffling,

1236
00:51:41,200 --> 00:51:43,920
smaller tweaks, and better

1237
00:51:42,240 --> 00:51:46,480
regularization.

1238
00:51:43,920 --> 00:51:48,559
Second was my hardware. My laptop

1239
00:51:46,480 --> 00:51:51,119
trained very slowly. It was like

1240
00:51:48,559 --> 00:51:53,359
watching paint dry. Uh I eventually had

1241
00:51:51,119 --> 00:51:55,760
to upgrade to my dad's um better

1242
00:51:53,359 --> 00:51:57,119
computer to run my program, which meant

1243
00:51:55,760 --> 00:51:59,119
that I could only train it when he was

1244
00:51:57,119 --> 00:52:01,040
sleeping.

1245
00:51:59,119 --> 00:52:05,359
So, my sleep schedule during that time

1246
00:52:01,040 --> 00:52:08,720
was horrendous. So bad. 3:00 a.m. And my

1247
00:52:05,359 --> 00:52:11,200
personal favorite, I misspelled bias.

1248
00:52:08,720 --> 00:52:15,119
Um,

1249
00:52:11,200 --> 00:52:17,359
two entire weeks of panic rediving

1250
00:52:15,119 --> 00:52:19,200
equations only to re realize that it was

1251
00:52:17,359 --> 00:52:21,599
a typo.

1252
00:52:19,200 --> 00:52:23,359
The scariest bug in my entire network

1253
00:52:21,599 --> 00:52:25,440
wasn't the math or the theory. It was a

1254
00:52:23,359 --> 00:52:27,680
single typo.

1255
00:52:25,440 --> 00:52:31,400
But overall, it was a fun experience. I

1256
00:52:27,680 --> 00:52:31,400
don't recommend it though.

1257
00:52:31,839 --> 00:52:38,160
Um, so that's Bonu and Numpai only C CNN

1258
00:52:36,079 --> 00:52:40,480
that can read handwriting at 95%

1259
00:52:38,160 --> 00:52:42,240
accuracy after 10 hours of training, a

1260
00:52:40,480 --> 00:52:44,730
lot of coffee and two weeks lost to a

1261
00:52:42,240 --> 00:52:50,800
typo. Thank you.

1262
00:52:44,730 --> 00:52:53,520
[Applause]

1263
00:52:50,800 --> 00:52:55,440
That's amazing. What did your dad think

1264
00:52:53,520 --> 00:52:58,559
about having come up there? Come on.

1265
00:52:55,440 --> 00:53:00,160
Come go. Yes. Yeah. What did your dad

1266
00:52:58,559 --> 00:53:02,000
think about you just stealing his laptop

1267
00:53:00,160 --> 00:53:04,079
for for This is This is for work, I

1268
00:53:02,000 --> 00:53:06,079
swear. It's important learning.

1269
00:53:04,079 --> 00:53:07,119
He did not mind, actually, which is

1270
00:53:06,079 --> 00:53:08,720
surprising.

1271
00:53:07,119 --> 00:53:11,520
Possibly cuz you only used it while he

1272
00:53:08,720 --> 00:53:15,200
was asleep.

1273
00:53:11,520 --> 00:53:17,090
Amazing. Uh, hang. Yeah, another massive

1274
00:53:15,200 --> 00:53:22,389
round. Give me one of

1275
00:53:17,090 --> 00:53:22,389
[Applause]

1276
00:53:24,160 --> 00:53:28,559
up next

1277
00:53:26,160 --> 00:53:32,079
we have

1278
00:53:28,559 --> 00:53:35,359
Shri. Yes, I can read my table that I

1279
00:53:32,079 --> 00:53:39,440
did. Um, Shri is going to be talking

1280
00:53:35,359 --> 00:53:42,720
about this. Uh, Shri has been

1281
00:53:39,440 --> 00:53:45,440
programming for eight years.

1282
00:53:42,720 --> 00:53:48,800
Her favorite emoji

1283
00:53:45,440 --> 00:53:51,839
is the crying face emoji.

1284
00:53:48,800 --> 00:53:53,760
A uh and see I hope that's not

1285
00:53:51,839 --> 00:53:56,800
reflective of her experience in doing

1286
00:53:53,760 --> 00:53:58,720
this. I fear it. No. Um uh and she

1287
00:53:56,800 --> 00:54:00,960
recently found out that the very first

1288
00:53:58,720 --> 00:54:04,559
computer bug was in fact literally a

1289
00:54:00,960 --> 00:54:07,839
bug. Uh, in 1947, uh, Rear Admiral Grace

1290
00:54:04,559 --> 00:54:11,280
Hopper, legend, uh, and her team found a

1291
00:54:07,839 --> 00:54:13,359
moth stuck in the relays of a of the

1292
00:54:11,280 --> 00:54:16,559
Harvard MK2 computer, which was causing

1293
00:54:13,359 --> 00:54:18,960
the machines to malfunction, hence

1294
00:54:16,559 --> 00:54:22,079
debugging, just, you know, getting those

1295
00:54:18,960 --> 00:54:26,119
moths out. Uh, amazing, Shri, take it

1296
00:54:22,079 --> 00:54:26,119
away. Applause for Yeah.

1297
00:54:29,280 --> 00:54:34,240
So hi, my name is Shri and my project is

1298
00:54:32,319 --> 00:54:36,640
using CNN's for breast cancer

1299
00:54:34,240 --> 00:54:38,960
hisystologology detection.

1300
00:54:36,640 --> 00:54:41,599
So a little bit of background about me.

1301
00:54:38,960 --> 00:54:43,440
Um I'm a year 12 student in Berglin

1302
00:54:41,599 --> 00:54:45,040
Anglican School from Canra. I'm

1303
00:54:43,440 --> 00:54:47,119
currently studying a major in computer

1304
00:54:45,040 --> 00:54:49,119
science. And a fun fact about me is that

1305
00:54:47,119 --> 00:54:50,559
I love doing eye makeup. This was the

1306
00:54:49,119 --> 00:54:52,880
makeup I did when I dressed up as

1307
00:54:50,559 --> 00:54:55,040
Cleopatra for book week this year. And

1308
00:54:52,880 --> 00:54:58,720
yes, I did actually have this project

1309
00:54:55,040 --> 00:55:00,640
idea back in year 10.

1310
00:54:58,720 --> 00:55:02,480
So this year I finally got around to

1311
00:55:00,640 --> 00:55:04,559
image classification, which has been my

1312
00:55:02,480 --> 00:55:07,200
dream project since year 10. So finally

1313
00:55:04,559 --> 00:55:09,520
we're here.

1314
00:55:07,200 --> 00:55:12,000
So this project classifies breast tissue

1315
00:55:09,520 --> 00:55:14,400
biopsy images from the break data set

1316
00:55:12,000 --> 00:55:16,240
into either benign or malignant where

1317
00:55:14,400 --> 00:55:18,000
benign means there's no cancer present

1318
00:55:16,240 --> 00:55:20,480
and malignant means there is a cancer

1319
00:55:18,000 --> 00:55:22,800
present. So basically we're training the

1320
00:55:20,480 --> 00:55:26,599
computer to be a tiny pathologist just

1321
00:55:22,800 --> 00:55:26,599
without the copy addiction.

1322
00:55:27,440 --> 00:55:31,680
I used a convolutional neural network or

1323
00:55:29,680 --> 00:55:33,920
CNN which is perfect for image

1324
00:55:31,680 --> 00:55:35,920
classification. Stick with me and by the

1325
00:55:33,920 --> 00:55:38,160
end you'll know why it's basically magic

1326
00:55:35,920 --> 00:55:40,640
for images.

1327
00:55:38,160 --> 00:55:42,720
So now I'll briefly explain how a CNN

1328
00:55:40,640 --> 00:55:44,160
works using digit classification as an

1329
00:55:42,720 --> 00:55:46,640
example.

1330
00:55:44,160 --> 00:55:48,880
An image is just a grid of pixels with

1331
00:55:46,640 --> 00:55:52,559
each pixel's grayscale value fed into

1332
00:55:48,880 --> 00:55:57,040
the network as an input neuron.

1333
00:55:52,559 --> 00:56:00,079
For a 28x 28 pixel, that means 784 input

1334
00:55:57,040 --> 00:56:02,559
values. Sounds like a lot, right? Hidden

1335
00:56:00,079 --> 00:56:04,960
layers use filters to scan small patches

1336
00:56:02,559 --> 00:56:06,640
of pixels at a time for patterns. Kind

1337
00:56:04,960 --> 00:56:09,040
of like a mini Sherlock Holmes for

1338
00:56:06,640 --> 00:56:11,839
pixels.

1339
00:56:09,040 --> 00:56:13,920
A 3x3 filter multiply multiplies each

1340
00:56:11,839 --> 00:56:16,000
pixel by a weight and sums these dot

1341
00:56:13,920 --> 00:56:17,920
products. Basically, the network is

1342
00:56:16,000 --> 00:56:21,359
trying to figure out if this tiny patch

1343
00:56:17,920 --> 00:56:24,000
is part of a three or a seven.

1344
00:56:21,359 --> 00:56:26,319
Then we add a bias, apply an activation

1345
00:56:24,000 --> 00:56:28,160
function like relu or sigmoid, and boom,

1346
00:56:26,319 --> 00:56:30,799
we get a neuron value for the next

1347
00:56:28,160 --> 00:56:32,640
layer. This number tells the network how

1348
00:56:30,799 --> 00:56:35,440
strongly that patch matches the pattern

1349
00:56:32,640 --> 00:56:37,920
that it's looking for.

1350
00:56:35,440 --> 00:56:40,000
This process repeats through all hidden

1351
00:56:37,920 --> 00:56:42,799
layers using weights and dot products

1352
00:56:40,000 --> 00:56:45,119
from the previous layer.

1353
00:56:42,799 --> 00:56:47,440
As filters slide across the image, they

1354
00:56:45,119 --> 00:56:48,960
form feature maps. Pooling layers

1355
00:56:47,440 --> 00:56:50,480
condense these maps, keeping the

1356
00:56:48,960 --> 00:56:53,040
strongest signals and speeding up

1357
00:56:50,480 --> 00:56:55,040
training. Fully connected layers combine

1358
00:56:53,040 --> 00:56:57,119
everything for the final prediction.

1359
00:56:55,040 --> 00:56:59,359
Step by step, the network moves from raw

1360
00:56:57,119 --> 00:57:02,000
pixels to identifying edges to shapes

1361
00:56:59,359 --> 00:57:03,839
and then full patterns.

1362
00:57:02,000 --> 00:57:05,440
The same process applies to

1363
00:57:03,839 --> 00:57:07,760
hisystologology images, but features

1364
00:57:05,440 --> 00:57:10,160
learn but filters learn features like

1365
00:57:07,760 --> 00:57:12,319
cell shapes, textures, and staining.

1366
00:57:10,160 --> 00:57:14,400
With enough label data, the network

1367
00:57:12,319 --> 00:57:16,880
separate separates benign from malignant

1368
00:57:14,400 --> 00:57:18,559
tissue. This relies on a process driven

1369
00:57:16,880 --> 00:57:20,480
by the loss function, which is basically

1370
00:57:18,559 --> 00:57:22,400
the network's way of saying, "Oops, I

1371
00:57:20,480 --> 00:57:24,640
messed up."

1372
00:57:22,400 --> 00:57:27,119
At the start, weights are random, so

1373
00:57:24,640 --> 00:57:29,359
predictions are terrible. Think my first

1374
00:57:27,119 --> 00:57:31,359
cooking attempt terrible.

1375
00:57:29,359 --> 00:57:35,040
The loss function measures the error and

1376
00:57:31,359 --> 00:57:36,960
minimizing it slowly improves accuracy.

1377
00:57:35,040 --> 00:57:39,200
Gradient descent adjusts weights to

1378
00:57:36,960 --> 00:57:41,839
reduce loss by converging towards a

1379
00:57:39,200 --> 00:57:43,920
local minimum in the cost function. Back

1380
00:57:41,839 --> 00:57:46,160
propagation calculates which weights

1381
00:57:43,920 --> 00:57:48,559
caused errors and the optimizer updates

1382
00:57:46,160 --> 00:57:50,480
them. So in short the process is

1383
00:57:48,559 --> 00:57:52,160
predict, compare, adjust and then

1384
00:57:50,480 --> 00:57:54,720
repeat.

1385
00:57:52,160 --> 00:57:56,640
I use PyTorch which handles the math for

1386
00:57:54,720 --> 00:57:58,400
convolutions and back propagation

1387
00:57:56,640 --> 00:58:00,960
letting us focus on model design and

1388
00:57:58,400 --> 00:58:02,960
training. So without PyTorch, I'd still

1389
00:58:00,960 --> 00:58:07,640
be handwriting partial derivatives and

1390
00:58:02,960 --> 00:58:07,640
this talk would be in PyCon 2035.

1391
00:58:08,400 --> 00:58:12,960
During training, the model cycles

1392
00:58:10,400 --> 00:58:15,760
through data over many epochs, gradually

1393
00:58:12,960 --> 00:58:17,920
improving accuracy. And yes, many epochs

1394
00:58:15,760 --> 00:58:19,760
is secret code for 20 minutes of staring

1395
00:58:17,920 --> 00:58:22,079
at graphs on the screen, praying it

1396
00:58:19,760 --> 00:58:26,079
works.

1397
00:58:22,079 --> 00:58:27,760
My CNN integrates ResNet 18 and TinyBG.

1398
00:58:26,079 --> 00:58:30,559
As much as I'd like to claim I wrote

1399
00:58:27,760 --> 00:58:32,480
ResNet 18 from scratch, I didn't.

1400
00:58:30,559 --> 00:58:35,040
Pre-trained models are wonderful things,

1401
00:58:32,480 --> 00:58:37,520
especially for histologology. Most of my

1402
00:58:35,040 --> 00:58:39,280
work was integrating existing code for

1403
00:58:37,520 --> 00:58:41,760
pre-processing and importing these

1404
00:58:39,280 --> 00:58:45,440
models. And these were the tutorials I

1405
00:58:41,760 --> 00:58:48,400
used and code that I integrated.

1406
00:58:45,440 --> 00:58:50,319
ResNet 18 has 18 layers as the name

1407
00:58:48,400 --> 00:58:52,000
suggests and uses skip connections to

1408
00:58:50,319 --> 00:58:54,640
preserve information and speed up

1409
00:58:52,000 --> 00:58:56,559
learning. I've used ResNet as my feature

1410
00:58:54,640 --> 00:58:58,319
extractor to identify the patterns in

1411
00:58:56,559 --> 00:59:00,559
the image.

1412
00:58:58,319 --> 00:59:02,799
I've replaced its final fully connected

1413
00:59:00,559 --> 00:59:05,200
layer with a tiny BGG inspired

1414
00:59:02,799 --> 00:59:09,040
classifier head that flattens features,

1415
00:59:05,200 --> 00:59:11,520
applies RLU and outputs predictions.

1416
00:59:09,040 --> 00:59:13,680
And this is how I've integrated the two.

1417
00:59:11,520 --> 00:59:16,079
Basically, ResNet finds the features and

1418
00:59:13,680 --> 00:59:18,079
tiny VG makes the call. Teamwork makes

1419
00:59:16,079 --> 00:59:20,000
the dream work.

1420
00:59:18,079 --> 00:59:22,400
For simplicity, I worked with one

1421
00:59:20,000 --> 00:59:25,280
malignant and one benign subcategory at

1422
00:59:22,400 --> 00:59:29,520
40 times magnification.

1423
00:59:25,280 --> 00:59:34,079
An 8020 split gave 893 training images

1424
00:59:29,520 --> 00:59:36,160
and 224 testing images.

1425
00:59:34,079 --> 00:59:38,640
As seen in the graphs, augmentation of

1426
00:59:36,160 --> 00:59:41,680
data with applied rotations and flips

1427
00:59:38,640 --> 00:59:43,200
boosted testing accuracy to about 97%.

1428
00:59:41,680 --> 00:59:44,880
And yes, this is the part where I

1429
00:59:43,200 --> 00:59:47,880
quietly celebrated with a little happy

1430
00:59:44,880 --> 00:59:47,880
dance.

1431
00:59:48,559 --> 00:59:55,200
So here is project demo.

1432
00:59:52,799 --> 00:59:57,359
So here a random image from the testing

1433
00:59:55,200 --> 01:00:00,000
data set is selected and inputed into

1434
00:59:57,359 --> 01:00:01,920
the model. the prediction, true label,

1435
01:00:00,000 --> 01:00:03,760
and whether the whether the two matched

1436
01:00:01,920 --> 01:00:05,359
is displayed here. And we can continue

1437
01:00:03,760 --> 01:00:07,520
rerunning this as many times as we'd

1438
01:00:05,359 --> 01:00:09,440
like. Spoiler, sometimes it's wrong,

1439
01:00:07,520 --> 01:00:11,359
which is about 3% of times, and that's

1440
01:00:09,440 --> 01:00:14,760
okay. It gives us something to optimize

1441
01:00:11,359 --> 01:00:14,760
in the future.

1442
01:00:17,839 --> 01:00:22,720
So, in the hybrid CNN, I replaced only

1443
01:00:20,559 --> 01:00:25,839
the last layer of ResNet 18 with the

1444
01:00:22,720 --> 01:00:28,720
tiny BGG inspired classifier.

1445
01:00:25,839 --> 01:00:30,880
A tiny typo forgetting the colon made

1446
01:00:28,720 --> 01:00:32,880
the model select only the last layer

1447
01:00:30,880 --> 01:00:35,040
instead of all of the layers except the

1448
01:00:32,880 --> 01:00:36,880
final one. This meant that there was

1449
01:00:35,040 --> 01:00:39,760
actually no feature extraction going on

1450
01:00:36,880 --> 01:00:41,839
and accuracy dropped to 20%.

1451
01:00:39,760 --> 01:00:43,839
So after scanning for hours and

1452
01:00:41,839 --> 01:00:48,480
questioning my life decisions, I fixed

1453
01:00:43,839 --> 01:00:51,760
it and accuracy rose back up to 96%.

1454
01:00:48,480 --> 01:00:53,599
Okay, so this is actually a potential

1455
01:00:51,760 --> 01:00:55,520
issue, not something I actually face

1456
01:00:53,599 --> 01:00:57,839
because ResNet 18 is already pre-trained

1457
01:00:55,520 --> 01:01:00,079
model. But one of the biggest issues

1458
01:00:57,839 --> 01:01:01,920
with gradient descent is getting stuck

1459
01:01:00,079 --> 01:01:04,000
in local minima, especially in

1460
01:01:01,920 --> 01:01:06,000
relatively flat regions of the cost

1461
01:01:04,000 --> 01:01:08,160
curve instead of converging towards a

1462
01:01:06,000 --> 01:01:09,839
local minimum. The model thinks it's

1463
01:01:08,160 --> 01:01:12,240
done, but it hasn't actually found the

1464
01:01:09,839 --> 01:01:13,920
best solution yet. So I might encounter

1465
01:01:12,240 --> 01:01:16,079
this if I actually build ResNet from

1466
01:01:13,920 --> 01:01:18,079
scratch next semester instead of simp

1467
01:01:16,079 --> 01:01:20,000
simply importing it and using it and I

1468
01:01:18,079 --> 01:01:24,920
hope to explore solutions to such a

1469
01:01:20,000 --> 01:01:24,920
problem then. Thank you.

1470
01:01:26,799 --> 01:01:29,799
Amazing.

1471
01:01:29,839 --> 01:01:34,559
That was fantastic.

1472
01:01:32,880 --> 01:01:37,280
I hope you're all thinking up lots of

1473
01:01:34,559 --> 01:01:40,680
amazing questions.

1474
01:01:37,280 --> 01:01:40,680
I am

1475
01:01:48,640 --> 01:01:53,480
no.

1476
01:01:50,079 --> 01:01:53,480
Round of applause.

1477
01:01:53,920 --> 01:02:01,440
Thank you very much. Okay,

1478
01:01:56,880 --> 01:02:04,480
next up we have Leticia. Leticia is

1479
01:02:01,440 --> 01:02:07,680
going to be talking about a moodbased

1480
01:02:04,480 --> 01:02:10,319
playlist generator.

1481
01:02:07,680 --> 01:02:15,040
She's been programming for three years

1482
01:02:10,319 --> 01:02:19,960
and her favorite emoji is the

1483
01:02:15,040 --> 01:02:19,960
monle quizzical face.

1484
01:02:20,240 --> 01:02:25,119
Um, she recently learned about what deep

1485
01:02:23,119 --> 01:02:27,520
learning and neural networks are and

1486
01:02:25,119 --> 01:02:31,200
their applications,

1487
01:02:27,520 --> 01:02:33,920
which is pretty fantastic.

1488
01:02:31,200 --> 01:02:37,440
Beautiful.

1489
01:02:33,920 --> 01:02:37,440
Okay. Yes.

1490
01:02:42,559 --> 01:02:47,440
Um, good afternoon. My name is Leticia

1491
01:02:45,680 --> 01:02:49,040
and today I'm going to be presenting my

1492
01:02:47,440 --> 01:02:51,920
project, which is a moodbased playlist

1493
01:02:49,040 --> 01:02:53,839
generator. Some facts about me are that

1494
01:02:51,920 --> 01:02:56,000
I like to listen to lots of music in

1495
01:02:53,839 --> 01:02:58,160
different languages and genres. I

1496
01:02:56,000 --> 01:02:59,920
created this application to address my

1497
01:02:58,160 --> 01:03:03,599
personal issue with trying to find

1498
01:02:59,920 --> 01:03:06,000
varied songs that resonate with my mood.

1499
01:03:03,599 --> 01:03:07,920
Um, yes, I have studied IT and computer

1500
01:03:06,000 --> 01:03:10,000
science at school for almost four years,

1501
01:03:07,920 --> 01:03:12,079
and I'm currently doing a school-based

1502
01:03:10,000 --> 01:03:14,559
digital technologies apprenticeship at

1503
01:03:12,079 --> 01:03:17,200
DU and I may do a Bachelor of Science at

1504
01:03:14,559 --> 01:03:19,039
A&U next year.

1505
01:03:17,200 --> 01:03:20,720
Machine learning is creating models and

1506
01:03:19,039 --> 01:03:22,799
algorithms that enable computers to

1507
01:03:20,720 --> 01:03:24,160
learn from data automatically. Machine

1508
01:03:22,799 --> 01:03:25,920
learning models can make accurate

1509
01:03:24,160 --> 01:03:28,079
predictions and intelligent decisions

1510
01:03:25,920 --> 01:03:29,599
based on learned knowledge. I use a

1511
01:03:28,079 --> 01:03:33,039
supervised learning approach where all

1512
01:03:29,599 --> 01:03:35,039
data used to train the model is labeled.

1513
01:03:33,039 --> 01:03:37,280
The key libraries and models that my

1514
01:03:35,039 --> 01:03:39,359
program uses are TensorFlow, which is an

1515
01:03:37,280 --> 01:03:40,960
open- source machine learning library

1516
01:03:39,359 --> 01:03:43,359
with features for building and training

1517
01:03:40,960 --> 01:03:45,440
neural networks. Caris, which is

1518
01:03:43,359 --> 01:03:47,599
TensorFlow's API that provides a

1519
01:03:45,440 --> 01:03:49,839
user-friendly interface for building

1520
01:03:47,599 --> 01:03:52,240
network architectures.

1521
01:03:49,839 --> 01:03:54,079
Um, Open AI's Whisper, which is an

1522
01:03:52,240 --> 01:03:56,960
automatic speech recognition model

1523
01:03:54,079 --> 01:03:59,359
trained on multilingual data. It enables

1524
01:03:56,960 --> 01:04:01,200
transcription from 50 languages and

1525
01:03:59,359 --> 01:04:03,920
translation from those languages into

1526
01:04:01,200 --> 01:04:06,000
English. Labs is a model that can

1527
01:04:03,920 --> 01:04:09,200
produce dimensional sentence embeddings

1528
01:04:06,000 --> 01:04:11,280
for 109 languages. And lastly, Librosa

1529
01:04:09,200 --> 01:04:12,880
is a specialized Python library which

1530
01:04:11,280 --> 01:04:17,200
provides functions to manipulate and

1531
01:04:12,880 --> 01:04:20,079
extract key features from audio files.

1532
01:04:17,200 --> 01:04:21,440
So my program has two pipelines. One for

1533
01:04:20,079 --> 01:04:23,920
sound which is in the form of

1534
01:04:21,440 --> 01:04:26,799
spectrograms and one for lyrics. for the

1535
01:04:23,920 --> 01:04:29,440
spectrogram pipeline. MP3s are organized

1536
01:04:26,799 --> 01:04:31,119
into folders labeled by their moods in a

1537
01:04:29,440 --> 01:04:32,720
directory. They are converted into

1538
01:04:31,119 --> 01:04:34,880
spectrograms organized in a new

1539
01:04:32,720 --> 01:04:36,240
directory with a mirrored structure. The

1540
01:04:34,880 --> 01:04:38,000
classified spectrograms are used to

1541
01:04:36,240 --> 01:04:40,480
train a convolutional neural network

1542
01:04:38,000 --> 01:04:42,400
model and the CNN classifies the moods

1543
01:04:40,480 --> 01:04:44,559
of unclassified songs based on what it

1544
01:04:42,400 --> 01:04:46,160
learned from the training data. Lastly,

1545
01:04:44,559 --> 01:04:50,400
it generates playlists containing songs

1546
01:04:46,160 --> 01:04:53,200
classified at the user's desired mood.

1547
01:04:50,400 --> 01:04:55,039
So I use Librosa to process MP3 audio

1548
01:04:53,200 --> 01:04:57,440
data to create and display MEL

1549
01:04:55,039 --> 01:04:59,039
spectrograms. The spectrograms are

1550
01:04:57,440 --> 01:05:00,799
converted to a decibel scale which is

1551
01:04:59,039 --> 01:05:02,799
logarithmic. This makes it more

1552
01:05:00,799 --> 01:05:05,440
perceptually accurate as humans perceive

1553
01:05:02,799 --> 01:05:07,520
volume changes logarithmically. Neural

1554
01:05:05,440 --> 01:05:10,000
networks also train better with well

1555
01:05:07,520 --> 01:05:12,400
distributed data and logarithmic scaling

1556
01:05:10,000 --> 01:05:14,240
normalizes spectral input. This

1557
01:05:12,400 --> 01:05:16,079
transforms an audio classification

1558
01:05:14,240 --> 01:05:19,760
problem into an image classification

1559
01:05:16,079 --> 01:05:22,000
task which CNN's excel at.

1560
01:05:19,760 --> 01:05:24,400
For the data generation, the code

1561
01:05:22,000 --> 01:05:26,480
expects a directory structure where MP3s

1562
01:05:24,400 --> 01:05:28,799
are organized into subfolders labeled by

1563
01:05:26,480 --> 01:05:31,200
mood so classifications can be

1564
01:05:28,799 --> 01:05:32,640
automatically inferred from the program.

1565
01:05:31,200 --> 01:05:34,400
The outputed spectrograms are

1566
01:05:32,640 --> 01:05:37,280
automatically organized into a mirrored

1567
01:05:34,400 --> 01:05:39,680
directory for training. I implement an

1568
01:05:37,280 --> 01:05:41,520
80 to 20 training validation data split

1569
01:05:39,680 --> 01:05:44,319
to evaluate how the model would perform

1570
01:05:41,520 --> 01:05:46,400
on unclassified data. The classification

1571
01:05:44,319 --> 01:05:48,240
method I chose was categorical rather

1572
01:05:46,400 --> 01:05:51,200
than dimensional which I found more

1573
01:05:48,240 --> 01:05:52,799
achievable in the time frame I had.

1574
01:05:51,200 --> 01:05:54,960
Now I will talk about the training and

1575
01:05:52,799 --> 01:05:56,559
model creation. But before I get into

1576
01:05:54,960 --> 01:05:59,680
the codes I will talk about the model I

1577
01:05:56,559 --> 01:06:01,440
use which is a CNN. CNN's are directed

1578
01:05:59,680 --> 01:06:03,119
to cyclic graphs where each node

1579
01:06:01,440 --> 01:06:05,920
computes a function of its inputs and

1580
01:06:03,119 --> 01:06:08,000
passes the result to its successes. This

1581
01:06:05,920 --> 01:06:10,240
works as hierarchical feature detectors

1582
01:06:08,000 --> 01:06:13,440
which is why CNN's are good for computer

1583
01:06:10,240 --> 01:06:15,359
vision tasks like image classification.

1584
01:06:13,440 --> 01:06:17,760
Um, a convolution is where a

1585
01:06:15,359 --> 01:06:20,240
matrix/filter slides over input to

1586
01:06:17,760 --> 01:06:21,680
generate a feature map. Each filter has

1587
01:06:20,240 --> 01:06:24,000
weights and biases which are shared

1588
01:06:21,680 --> 01:06:25,680
across all neurons that use it, reducing

1589
01:06:24,000 --> 01:06:28,079
the number of trainable parameters

1590
01:06:25,680 --> 01:06:30,319
significantly. And weights are how much

1591
01:06:28,079 --> 01:06:32,799
an input influences an output and the

1592
01:06:30,319 --> 01:06:34,720
weight sharing enables features filters

1593
01:06:32,799 --> 01:06:37,119
to detect features in the input

1594
01:06:34,720 --> 01:06:38,640
regardless of their location. Many

1595
01:06:37,119 --> 01:06:41,280
filters can be generated to capture

1596
01:06:38,640 --> 01:06:43,280
unique features from the input.

1597
01:06:41,280 --> 01:06:45,599
Pooling is a function that generates

1598
01:06:43,280 --> 01:06:47,839
smaller output vectors of the input. I

1599
01:06:45,599 --> 01:06:49,520
used max pooling which takes the maximum

1600
01:06:47,839 --> 01:06:51,599
value of the feature map and is good at

1601
01:06:49,520 --> 01:06:53,839
detecting the most prominent features in

1602
01:06:51,599 --> 01:06:55,680
the input. Pooling increases

1603
01:06:53,839 --> 01:06:57,760
computational efficiency and reduces

1604
01:06:55,680 --> 01:07:00,480
overfitting.

1605
01:06:57,760 --> 01:07:02,240
Um, ReLU is a mathematical function

1606
01:07:00,480 --> 01:07:04,559
applied to neuron outputs that

1607
01:07:02,240 --> 01:07:06,000
determines activation. With ReLU,

1608
01:07:04,559 --> 01:07:08,400
positive inputs are passed through

1609
01:07:06,000 --> 01:07:10,079
unchanged otherwise zero is outputed.

1610
01:07:08,400 --> 01:07:11,920
This creates representations where only

1611
01:07:10,079 --> 01:07:13,359
the most salient features activate,

1612
01:07:11,920 --> 01:07:16,079
ignoring weaker features that can

1613
01:07:13,359 --> 01:07:17,839
complicate decision- making. Lastly,

1614
01:07:16,079 --> 01:07:19,920
softmax is an activation function

1615
01:07:17,839 --> 01:07:21,680
applied to the network's last layer and

1616
01:07:19,920 --> 01:07:25,839
enables input classification by

1617
01:07:21,680 --> 01:07:28,559
assigning probabilities to each class.

1618
01:07:25,839 --> 01:07:30,400
So my model has a basic CNN architecture

1619
01:07:28,559 --> 01:07:34,000
optimized for image classification with

1620
01:07:30,400 --> 01:07:36,960
the components explained before.

1621
01:07:34,000 --> 01:07:38,559
Now for the lyric pipeline, MP3 files

1622
01:07:36,960 --> 01:07:40,559
from the original directory are

1623
01:07:38,559 --> 01:07:42,880
transcribed, translated to English if

1624
01:07:40,559 --> 01:07:45,359
necessary, and saved as text files in a

1625
01:07:42,880 --> 01:07:47,599
mirrored directory. Mood lyric pairs are

1626
01:07:45,359 --> 01:07:49,440
compiled into a CSV file, and lyrics in

1627
01:07:47,599 --> 01:07:51,440
the CSV are converted into sentence

1628
01:07:49,440 --> 01:07:53,359
embeddings used to train a deep neural

1629
01:07:51,440 --> 01:07:56,240
network.

1630
01:07:53,359 --> 01:07:58,319
For the MP3 to text conversion, OpenAI's

1631
01:07:56,240 --> 01:08:00,319
whisper model is loaded and used for

1632
01:07:58,319 --> 01:08:02,799
transcription and translation of MP3

1633
01:08:00,319 --> 01:08:04,720
files. An empty list is initialized to

1634
01:08:02,799 --> 01:08:07,680
store mood lyric pairs to be at the end

1635
01:08:04,720 --> 01:08:09,440
compiled into a CSV. Some issues I

1636
01:08:07,680 --> 01:08:11,280
experienced is that Whisper can be

1637
01:08:09,440 --> 01:08:12,960
resource intensive and takes a few

1638
01:08:11,280 --> 01:08:15,520
minutes to transcribe and translate a

1639
01:08:12,960 --> 01:08:17,759
single song and it often mistranscribes

1640
01:08:15,520 --> 01:08:19,520
or fails with noisy audio and is better

1641
01:08:17,759 --> 01:08:21,199
at translating some languages over

1642
01:08:19,520 --> 01:08:23,120
others.

1643
01:08:21,199 --> 01:08:26,120
Here's some example of a strange

1644
01:08:23,120 --> 01:08:26,120
translation.

1645
01:08:27,279 --> 01:08:32,000
Um the model used for this pipeline is a

1646
01:08:29,520 --> 01:08:33,920
transformer which is a neural network

1647
01:08:32,000 --> 01:08:35,199
architecture that processes text by

1648
01:08:33,920 --> 01:08:37,600
converting it into numerical

1649
01:08:35,199 --> 01:08:39,759
representations called tokens. Each

1650
01:08:37,600 --> 01:08:41,520
token is mapped to a vector using a word

1651
01:08:39,759 --> 01:08:43,120
embedding table allowing the model to

1652
01:08:41,520 --> 01:08:45,920
understand the meaning and context of

1653
01:08:43,120 --> 01:08:47,679
words. My transformer uses the lab's

1654
01:08:45,920 --> 01:08:49,600
model to produce sentence embeddings on

1655
01:08:47,679 --> 01:08:51,279
the lyrics. This means the model doesn't

1656
01:08:49,600 --> 01:08:53,120
have to learn from raw text but more

1657
01:08:51,279 --> 01:08:56,159
informative embeddings and the model can

1658
01:08:53,120 --> 01:08:58,560
map meaning to mood.

1659
01:08:56,159 --> 01:09:00,159
And the last step is classification. And

1660
01:08:58,560 --> 01:09:02,960
the above images show the general

1661
01:09:00,159 --> 01:09:04,960
classification pipeline where the system

1662
01:09:02,960 --> 01:09:07,440
initializes a global playlist dictionary

1663
01:09:04,960 --> 01:09:09,839
with empty lists for each mood category.

1664
01:09:07,440 --> 01:09:11,600
For each MP3 file, a spectrogram image

1665
01:09:09,839 --> 01:09:13,600
or text file is generated through the

1666
01:09:11,600 --> 01:09:15,359
affforementioned processing techniques

1667
01:09:13,600 --> 01:09:17,759
and inputed to the trained models for

1668
01:09:15,359 --> 01:09:19,359
analysis. Each processed song is

1669
01:09:17,759 --> 01:09:21,520
automatically sorted into its mood

1670
01:09:19,359 --> 01:09:23,040
category within the playlist dictionary,

1671
01:09:21,520 --> 01:09:26,080
creating an organized collection of

1672
01:09:23,040 --> 01:09:28,880
playlists defined by mood.

1673
01:09:26,080 --> 01:09:31,880
And now I can show the product of my

1674
01:09:28,880 --> 01:09:31,880
code.

1675
01:09:37,359 --> 01:09:41,319
Um, does anyone know?

1676
01:09:51,279 --> 01:09:56,239
We're building suspense for the demo.

1677
01:10:02,000 --> 01:10:06,280
Yeah, I kind of want to show this one.

1678
01:10:13,679 --> 01:10:20,040
The quietest foot drum roll

1679
01:10:17,040 --> 01:10:20,040
spreads.

1680
01:10:21,040 --> 01:10:25,199
This is how you get your steps up.

1681
01:10:25,679 --> 01:10:28,679
Switch.

1682
01:10:44,719 --> 01:10:49,719
You didn't realize this was a fitness

1683
01:10:46,480 --> 01:10:49,719
boot camp.

1684
01:10:50,320 --> 01:10:52,800
Oh,

1685
01:10:51,199 --> 01:10:58,280
dope.

1686
01:10:52,800 --> 01:10:58,280
B. Yeah. Okay. Burpees now. Hey.

1687
01:10:59,520 --> 01:11:03,840
So, you can see there's two types you

1688
01:11:01,520 --> 01:11:06,640
can choose from. Either sound or lyrics

1689
01:11:03,840 --> 01:11:08,159
and it will generate different playlists

1690
01:11:06,640 --> 01:11:12,400
probably because it uses different

1691
01:11:08,159 --> 01:11:14,159
models. And then there's the moods here.

1692
01:11:12,400 --> 01:11:16,640
Some playlists don't have as many

1693
01:11:14,159 --> 01:11:19,440
because I guess the model was

1694
01:11:16,640 --> 01:11:22,920
overfitting or something. And yeah,

1695
01:11:19,440 --> 01:11:22,920
that's about it.

1696
01:11:25,679 --> 01:11:30,199
A massive round of applause.

1697
01:11:30,719 --> 01:11:35,199
The demo gods are smiling on us. They

1698
01:11:33,280 --> 01:11:36,719
are rewarding.

1699
01:11:35,199 --> 01:11:38,000
I was going to say they're rewarding our

1700
01:11:36,719 --> 01:11:42,040
patients, but really they're they're

1701
01:11:38,000 --> 01:11:42,040
rewarding the students efforts.

1702
01:11:44,719 --> 01:11:52,800
High five. Yes. Amazing.

1703
01:11:48,960 --> 01:11:54,480
One more massive round of applause.

1704
01:11:52,800 --> 01:11:56,800
Okay,

1705
01:11:54,480 --> 01:11:58,800
we have two talks left. I know you're

1706
01:11:56,800 --> 01:12:01,120
all tired, all of that stomping your

1707
01:11:58,800 --> 01:12:04,960
feet. It's okay.

1708
01:12:01,120 --> 01:12:06,800
It's all right. Uh, up next we have

1709
01:12:04,960 --> 01:12:10,000
Kevin.

1710
01:12:06,800 --> 01:12:13,679
Kevin is going to talk about Python

1711
01:12:10,000 --> 01:12:15,600
Flask as a backend framework. Um, Kevin

1712
01:12:13,679 --> 01:12:18,159
was first introduced to programming in

1713
01:12:15,600 --> 01:12:21,360
year four, eight years ago for first

1714
01:12:18,159 --> 01:12:25,199
Lego League. Uh, Kevin's favorite emoji

1715
01:12:21,360 --> 01:12:28,239
is eyes.

1716
01:12:25,199 --> 01:12:30,239
Shifty eyes. Uh, and Kevin recently

1717
01:12:28,239 --> 01:12:32,800
learned about MCP servers, which help

1718
01:12:30,239 --> 01:12:35,760
standardize how LM receive context from

1719
01:12:32,800 --> 01:12:38,159
external sources. and he's been working

1720
01:12:35,760 --> 01:12:41,040
on a project that uses XMCP, the MCP

1721
01:12:38,159 --> 01:12:43,600
framework, to develop a custom MCP for

1722
01:12:41,040 --> 01:12:45,280
your app, the link to which you'll drop

1723
01:12:43,600 --> 01:12:50,199
in the Discord, right?

1724
01:12:45,280 --> 01:12:50,199
Yeah. Massive round of applause.

1725
01:12:52,320 --> 01:12:56,880
Hello. Um, I'm here to talk about Flask

1726
01:12:54,560 --> 01:12:58,560
as a fashion framework and um, recently

1727
01:12:56,880 --> 01:13:00,000
I was tasked by my teacher to create a

1728
01:12:58,560 --> 01:13:01,280
web app and normally that'll be

1729
01:13:00,000 --> 01:13:03,280
straightforward. I would use like a

1730
01:13:01,280 --> 01:13:05,040
JavaScript framework like React, but

1731
01:13:03,280 --> 01:13:08,159
unfortunately there was one exception. I

1732
01:13:05,040 --> 01:13:10,560
had to use Python. And um so I decided

1733
01:13:08,159 --> 01:13:13,360
to go with a hybrid approach would which

1734
01:13:10,560 --> 01:13:15,199
would include using both. And so I'll

1735
01:13:13,360 --> 01:13:18,480
use Flask as my backend framework and

1736
01:13:15,199 --> 01:13:20,560
React as my front-end framework. Um so

1737
01:13:18,480 --> 01:13:22,239
there was just one problem. It was I'd

1738
01:13:20,560 --> 01:13:24,320
never built an app before with two

1739
01:13:22,239 --> 01:13:28,640
different back end. Oh, with one front

1740
01:13:24,320 --> 01:13:32,480
end and one separate back end. Um,

1741
01:13:28,640 --> 01:13:34,400
so sorry. Um, yeah. So, my app that I've

1742
01:13:32,480 --> 01:13:36,640
designed is Fineman and it basically

1743
01:13:34,400 --> 01:13:39,360
unifies workflows for students and it

1744
01:13:36,640 --> 01:13:41,679
interlinks tasks, notes as well as time

1745
01:13:39,360 --> 01:13:43,679
study sessions into one timeline so you

1746
01:13:41,679 --> 01:13:45,840
know exactly what you're working on,

1747
01:13:43,679 --> 01:13:47,440
when, and why. And this helps you to

1748
01:13:45,840 --> 01:13:49,040
further optimize your productivity

1749
01:13:47,440 --> 01:13:51,360
continuously.

1750
01:13:49,040 --> 01:13:53,920
Um, and this means that the app is

1751
01:13:51,360 --> 01:13:55,679
storing sensitive data such as personal

1752
01:13:53,920 --> 01:13:57,920
email addresses as well as personal

1753
01:13:55,679 --> 01:14:00,159
notes and reflections. So if an attacker

1754
01:13:57,920 --> 01:14:02,560
gained access to the system, it would

1755
01:14:00,159 --> 01:14:04,719
compromise the integrity of the app. And

1756
01:14:02,560 --> 01:14:07,120
this makes the browser server boundary

1757
01:14:04,719 --> 01:14:09,120
very important and critical to the app's

1758
01:14:07,120 --> 01:14:11,600
operation.

1759
01:14:09,120 --> 01:14:14,000
Um so initially since I had no previous

1760
01:14:11,600 --> 01:14:16,880
experience with blast I decided to put

1761
01:14:14,000 --> 01:14:19,199
all my validation on the client side and

1762
01:14:16,880 --> 01:14:21,440
unfortunately this is not very safe um

1763
01:14:19,199 --> 01:14:24,560
as it meant I had zero validation on the

1764
01:14:21,440 --> 01:14:26,880
server side. So as you can tell um this

1765
01:14:24,560 --> 01:14:28,640
could this meant that you could bypass

1766
01:14:26,880 --> 01:14:32,080
um all the security measures on the

1767
01:14:28,640 --> 01:14:35,280
client by making requests via like tools

1768
01:14:32,080 --> 01:14:37,600
such as Postmen and like the cult. And

1769
01:14:35,280 --> 01:14:40,320
um this meant that you could entirely

1770
01:14:37,600 --> 01:14:42,480
bypass all the chats and send all the

1771
01:14:40,320 --> 01:14:46,159
SQL injections or whatever code you want

1772
01:14:42,480 --> 01:14:47,920
directly to the server. Um now frankly

1773
01:14:46,159 --> 01:14:51,440
um I wasn't the only one who made this

1774
01:14:47,920 --> 01:14:53,040
mistake as um a couple of all um well

1775
01:14:51,440 --> 01:14:55,840
you've heard of British Airlines as well

1776
01:14:53,040 --> 01:14:59,440
as Uber. They've been fined and they've

1777
01:14:55,840 --> 01:15:02,239
um had significant reputational damage

1778
01:14:59,440 --> 01:15:05,360
as well as uh financial and legal uh

1779
01:15:02,239 --> 01:15:08,400
implications due to as a result of these

1780
01:15:05,360 --> 01:15:11,120
um security concerns.

1781
01:15:08,400 --> 01:15:13,199
Um so with that in mind, I decided to go

1782
01:15:11,120 --> 01:15:16,400
with a separate approach where I would

1783
01:15:13,199 --> 01:15:18,560
have an independent check on the server.

1784
01:15:16,400 --> 01:15:20,640
And so as you can see in the blue, um

1785
01:15:18,560 --> 01:15:22,880
there's an additional barrier for the

1786
01:15:20,640 --> 01:15:26,480
backend validation.

1787
01:15:22,880 --> 01:15:29,520
And this just processes and um verifies

1788
01:15:26,480 --> 01:15:32,000
inputs uh when the server receives uh

1789
01:15:29,520 --> 01:15:33,920
additional contacts. So um as you can

1790
01:15:32,000 --> 01:15:36,159
see I used a library called Cberus and

1791
01:15:33,920 --> 01:15:38,719
it has a validator which you can define

1792
01:15:36,159 --> 01:15:42,560
a schema and based off the schema it can

1793
01:15:38,719 --> 01:15:45,040
reject or um take in uh an input and

1794
01:15:42,560 --> 01:15:48,320
this runs regardless whether through the

1795
01:15:45,040 --> 01:15:50,159
UI or through an add an external API

1796
01:15:48,320 --> 01:15:54,000
request.

1797
01:15:50,159 --> 01:15:56,640
Um how oh and I also implemented cause

1798
01:15:54,000 --> 01:15:59,760
so um this aims to address cross origin

1799
01:15:56,640 --> 01:16:03,280
attacks. So these result from external

1800
01:15:59,760 --> 01:16:07,520
attacks from other sites where um you

1801
01:16:03,280 --> 01:16:09,760
could send uh malicious data um which

1802
01:16:07,520 --> 01:16:12,320
bypasses the cookie straight into the

1803
01:16:09,760 --> 01:16:14,239
server. So what causes it's a browser

1804
01:16:12,320 --> 01:16:16,159
rule that controls which websites and

1805
01:16:14,239 --> 01:16:18,560
addresses are able to read the response

1806
01:16:16,159 --> 01:16:20,719
from my back end. Um so as you can see

1807
01:16:18,560 --> 01:16:23,440
on the right with the Python code um

1808
01:16:20,719 --> 01:16:25,280
I've specifically defined the origins.

1809
01:16:23,440 --> 01:16:28,320
So I have the three main domains as well

1810
01:16:25,280 --> 01:16:30,719
as the local host server um to which are

1811
01:16:28,320 --> 01:16:34,080
the only domains that can take in and

1812
01:16:30,719 --> 01:16:37,280
read the requests from the back end.

1813
01:16:34,080 --> 01:16:38,960
Um however there is one downside um

1814
01:16:37,280 --> 01:16:41,520
cause does not stop the browser from

1815
01:16:38,960 --> 01:16:45,360
submitting HTML forms across sites. So,

1816
01:16:41,520 --> 01:16:48,159
this could expose uh my site to um

1817
01:16:45,360 --> 01:16:51,360
cross-sight forgery uh cross-sight

1818
01:16:48,159 --> 01:16:53,199
request forgery attacks. And this occurs

1819
01:16:51,360 --> 01:16:55,040
when a different site attempts to trick

1820
01:16:53,199 --> 01:16:57,600
the users's browser into sending a valid

1821
01:16:55,040 --> 01:17:00,080
cookie um which goes straight into the

1822
01:16:57,600 --> 01:17:03,280
back end. And since it uses the cookie

1823
01:17:00,080 --> 01:17:05,520
from the browser, uh it attaches all the

1824
01:17:03,280 --> 01:17:08,000
user data and basically is able to

1825
01:17:05,520 --> 01:17:10,000
effectively impersonate the user. And

1826
01:17:08,000 --> 01:17:12,239
this makes the user assume that it was

1827
01:17:10,000 --> 01:17:16,320
indeed the user rather than the militia

1828
01:17:12,239 --> 01:17:18,960
site. Um so as you can see um with a

1829
01:17:16,320 --> 01:17:22,239
simple demo on the right you can see a

1830
01:17:18,960 --> 01:17:24,000
simple form could basically like steal

1831
01:17:22,239 --> 01:17:28,239
your cookies from another site and

1832
01:17:24,000 --> 01:17:31,280
impersonate your um session to my site.

1833
01:17:28,239 --> 01:17:34,080
And again, this could lead to um attacks

1834
01:17:31,280 --> 01:17:36,640
such as like stealing information and as

1835
01:17:34,080 --> 01:17:40,400
well as deleting um personal details

1836
01:17:36,640 --> 01:17:43,040
which could lead to significant um in

1837
01:17:40,400 --> 01:17:47,120
data concerns.

1838
01:17:43,040 --> 01:17:50,080
So I moved on to my final iteration. Um

1839
01:17:47,120 --> 01:17:52,880
and this one I I implemented um

1840
01:17:50,080 --> 01:17:55,280
cross-ite refraudy request forgery

1841
01:17:52,880 --> 01:17:57,280
tokens. So this is a randomly generated

1842
01:17:55,280 --> 01:18:00,080
value which is issued by the server and

1843
01:17:57,280 --> 01:18:01,920
it's stored in the session and um this

1844
01:18:00,080 --> 01:18:04,719
makes it so that every request to the

1845
01:18:01,920 --> 01:18:06,800
API um must include the token in the

1846
01:18:04,719 --> 01:18:10,880
header or else it would return with a

1847
01:18:06,800 --> 01:18:13,760
400 403 error which is a forbidden um

1848
01:18:10,880 --> 01:18:15,280
error code and this made sure this

1849
01:18:13,760 --> 01:18:17,440
ensures that the back end and framework

1850
01:18:15,280 --> 01:18:19,199
uh back end and front end are able to

1851
01:18:17,440 --> 01:18:21,679
effectively and safely communicate

1852
01:18:19,199 --> 01:18:24,320
without issues.

1853
01:18:21,679 --> 01:18:26,640
Um well, was it was it worth it in the

1854
01:18:24,320 --> 01:18:29,440
end? Well, unfortunately, it turns out

1855
01:18:26,640 --> 01:18:31,199
not really because um there are

1856
01:18:29,440 --> 01:18:33,199
alternatives such as Django and Fast

1857
01:18:31,199 --> 01:18:35,600
API, which are also Python-based

1858
01:18:33,199 --> 01:18:38,000
frameworks that offer all of the

1859
01:18:35,600 --> 01:18:39,760
security implementations out of the box,

1860
01:18:38,000 --> 01:18:42,080
which means you wouldn't necessarily

1861
01:18:39,760 --> 01:18:43,840
have to implement this yourself. And as

1862
01:18:42,080 --> 01:18:46,719
you can see, they offer more than just

1863
01:18:43,840 --> 01:18:50,159
this. And they provide other benefits on

1864
01:18:46,719 --> 01:18:52,080
top of um the security measures.

1865
01:18:50,159 --> 01:18:54,640
But uh despite this, it was actually a

1866
01:18:52,080 --> 01:18:56,480
great learning experience um as it

1867
01:18:54,640 --> 01:18:59,120
provided me with a really in-depth and

1868
01:18:56,480 --> 01:19:01,199
foundational outlook on how security

1869
01:18:59,120 --> 01:19:03,280
works at a fundamental level. And

1870
01:19:01,199 --> 01:19:05,040
building it in FL in a FL environment

1871
01:19:03,280 --> 01:19:06,719
required me to make explicit and

1872
01:19:05,040 --> 01:19:09,520
intentional choices which gave me more

1873
01:19:06,719 --> 01:19:11,520
control over the intricacies of security

1874
01:19:09,520 --> 01:19:13,760
decisions.

1875
01:19:11,520 --> 01:19:15,600
Yeah. And um you can actually view the

1876
01:19:13,760 --> 01:19:19,360
site right now. Um the domain's on the

1877
01:19:15,600 --> 01:19:22,320
bottom right. It's fine. Um, and I'm

1878
01:19:19,360 --> 01:19:24,239
constantly making updates. Yeah. Um, as

1879
01:19:22,320 --> 01:19:27,280
mentioned earlier on, I'm working on

1880
01:19:24,239 --> 01:19:29,840
like adding MCP servers to the um, back

1881
01:19:27,280 --> 01:19:31,650
end so it works better with AI. And

1882
01:19:29,840 --> 01:19:38,000
yeah, thanks.

1883
01:19:31,650 --> 01:19:40,159
[Applause]

1884
01:19:38,000 --> 01:19:42,560
Amazing.

1885
01:19:40,159 --> 01:19:44,960
All right, you lot. Hang on. Hang on. I

1886
01:19:42,560 --> 01:19:46,320
need to high five you on the way out. I

1887
01:19:44,960 --> 01:19:50,679
have too many things in my hand. Yes,

1888
01:19:46,320 --> 01:19:50,679
thank you. But I need that as well.

1889
01:19:55,840 --> 01:20:01,400
Next up, we have a thing. So, we're

1890
01:19:57,840 --> 01:20:01,400
going to need a minute.

1891
01:20:11,520 --> 01:20:15,520
Now would be a fantastic time to think

1892
01:20:13,679 --> 01:20:17,199
of some questions and put them in the

1893
01:20:15,520 --> 01:20:20,199
Discord so you don't forget the

1894
01:20:17,199 --> 01:20:20,199
questions.

1895
01:20:20,239 --> 01:20:27,960
Um, oh, I've got so many people to

1896
01:20:22,880 --> 01:20:27,960
introduce. I should not be stalling. Um,

1897
01:20:28,000 --> 01:20:31,760
I should be stalling

1898
01:20:30,080 --> 01:20:35,280
with the proper stalling here. Hold

1899
01:20:31,760 --> 01:20:39,040
this. Thank you. Excellent. Okay, so

1900
01:20:35,280 --> 01:20:40,960
next up we have

1901
01:20:39,040 --> 01:20:43,600
Sigund.

1902
01:20:40,960 --> 01:20:47,440
This is Sigund.

1903
01:20:43,600 --> 01:20:48,960
Um, we have Alex, Haley, Lonnie, and

1904
01:20:47,440 --> 01:20:50,000
Ryland.

1905
01:20:48,960 --> 01:20:53,280
I'm gonna do that again. You're going to

1906
01:20:50,000 --> 01:20:56,000
wave at the right time, Alex. Hang on.

1907
01:20:53,280 --> 01:20:59,360
Yeah. Ah, there we go. Even nice.

1908
01:20:56,000 --> 01:21:03,280
Beautiful. Um, their favorite emoji

1909
01:20:59,360 --> 01:21:04,960
respectively are sunglasses colon

1910
01:21:03,280 --> 01:21:07,520
bracket bracket. I know it's not an

1911
01:21:04,960 --> 01:21:13,199
emoji, but I like it anyway. Hands

1912
01:21:07,520 --> 01:21:16,880
together. And gave me two options of

1913
01:21:13,199 --> 01:21:18,719
I eyes with no face. I don't know why I

1914
01:21:16,880 --> 01:21:22,640
closed my eyes but continued to have a

1915
01:21:18,719 --> 01:21:25,360
face then. I'm not really sure. um or

1916
01:21:22,640 --> 01:21:28,880
pro perhaps uh relevant kind of sad

1917
01:21:25,360 --> 01:21:31,360
looking um they recently they've they've

1918
01:21:28,880 --> 01:21:35,120
been programming a while eight years

1919
01:21:31,360 --> 01:21:37,120
since year seven four years uh and about

1920
01:21:35,120 --> 01:21:39,199
seven years ago.

1921
01:21:37,120 --> 01:21:42,080
Um they've learned that it's pronounced

1922
01:21:39,199 --> 01:21:45,040
GIF, not GIF.

1923
01:21:42,080 --> 01:21:47,040
Um

1924
01:21:45,040 --> 01:21:50,239
uh they've learned how neat algorithms

1925
01:21:47,040 --> 01:21:52,560
work. They've they funnily enough, it

1926
01:21:50,239 --> 01:21:56,719
was this question that taught me you can

1927
01:21:52,560 --> 01:22:00,560
combine emoji. Yeah. And they've learned

1928
01:21:56,719 --> 01:22:02,800
that writing the report is the least fun

1929
01:22:00,560 --> 01:22:05,360
bit of a coding project.

1930
01:22:02,800 --> 01:22:09,199
[Applause]

1931
01:22:05,360 --> 01:22:13,080
I'm gonna Yes. Round of applause.

1932
01:22:09,199 --> 01:22:13,080
I'm going to give you two microphones.

1933
01:22:13,120 --> 01:22:17,600
That is That's for Sigment.

1934
01:22:16,000 --> 01:22:20,600
I've been told this microphone is for

1935
01:22:17,600 --> 01:22:20,600
segment.

1936
01:22:22,400 --> 01:22:27,480
Uh good evening all. We are the team

1937
01:22:23,920 --> 01:22:27,480
behind Sigment.

1938
01:22:27,760 --> 01:22:33,040
So what is Sigman? Sigment is the

1939
01:22:30,560 --> 01:22:35,760
combines both software and hardware to

1940
01:22:33,040 --> 01:22:38,080
create a safe space where everyone can

1941
01:22:35,760 --> 01:22:40,239
share their troubles freely.

1942
01:22:38,080 --> 01:22:43,120
We built Sigund as a way to lower the

1943
01:22:40,239 --> 01:22:46,239
barriers to mental health care, a a

1944
01:22:43,120 --> 01:22:48,320
topic that touches us all.

1945
01:22:46,239 --> 01:22:49,840
Um so Sigman works as like a middleman

1946
01:22:48,320 --> 01:22:54,159
between

1947
01:22:49,840 --> 01:22:56,400
um the client and so you and him um

1948
01:22:54,159 --> 01:22:58,159
trying to

1949
01:22:56,400 --> 01:23:00,000
make make this way like to make therapy

1950
01:22:58,159 --> 01:23:01,920
more affordable, more accessible to

1951
01:23:00,000 --> 01:23:04,560
people, just make everything a lot

1952
01:23:01,920 --> 01:23:07,040
easier. So we have a little teaser just

1953
01:23:04,560 --> 01:23:10,360
for a start.

1954
01:23:07,040 --> 01:23:10,360
Go say hi.

1955
01:23:13,600 --> 01:23:18,679
Back on back.

1956
01:23:15,679 --> 01:23:18,679
Back.

1957
01:23:23,360 --> 01:23:26,360
Hi.

1958
01:23:30,960 --> 01:23:35,120
It is

1959
01:23:32,719 --> 01:23:37,040
Penny Prit. So great to hear from you.

1960
01:23:35,120 --> 01:23:39,520
How have you been doing since we last

1961
01:23:37,040 --> 01:23:41,760
spoke? I've been thinking about our

1962
01:23:39,520 --> 01:23:43,840
previous conversations and wanted to

1963
01:23:41,760 --> 01:23:46,840
check in on how things are going for

1964
01:23:43,840 --> 01:23:46,840
you.

1965
01:23:49,030 --> 01:23:52,560
[Applause]

1966
01:23:50,239 --> 01:23:54,080
Uh so, Sigment is made up of a bunch of

1967
01:23:52,560 --> 01:23:56,560
different components that all work

1968
01:23:54,080 --> 01:24:00,080
together in harmony to create the weird

1969
01:23:56,560 --> 01:24:04,480
thing you see in front of you. Um first

1970
01:24:00,080 --> 01:24:08,520
of all, uh I worked on the server and

1971
01:24:04,480 --> 01:24:08,520
the large language model.

1972
01:24:09,280 --> 01:24:15,360
Um, I worked on the GUI so it can be

1973
01:24:12,320 --> 01:24:19,960
more accessible for users.

1974
01:24:15,360 --> 01:24:19,960
I built the face and made it emotional.

1975
01:24:20,239 --> 01:24:24,840
And I did your your input and output.

1976
01:24:25,360 --> 01:24:30,560
Um, so the server is a Flask framework

1977
01:24:27,760 --> 01:24:32,239
with uh a SQL database as the back end

1978
01:24:30,560 --> 01:24:33,920
which allows users to communicate and

1979
01:24:32,239 --> 01:24:36,560
bring up previous history that they've

1980
01:24:33,920 --> 01:24:40,159
talked to with Sigman beforehand. Now

1981
01:24:36,560 --> 01:24:42,560
the LLM Sigman says that he is this

1982
01:24:40,159 --> 01:24:45,360
thing here. Uh it's very good, sounds

1983
01:24:42,560 --> 01:24:49,040
human. Unfortunately, it is all wrong.

1984
01:24:45,360 --> 01:24:50,480
Um Sigman is actually Llama 3.2, which

1985
01:24:49,040 --> 01:24:52,880
is a transformer model that's trained on

1986
01:24:50,480 --> 01:24:54,800
a very big corpus of text by Llama. It

1987
01:24:52,880 --> 01:24:58,719
makes the LLM not only responsive, but

1988
01:24:54,800 --> 01:25:00,560
also incredibly human. It's scary, but

1989
01:24:58,719 --> 01:25:02,639
it's really good for something like

1990
01:25:00,560 --> 01:25:04,800
talking to someone and maintaining that

1991
01:25:02,639 --> 01:25:07,199
connection.

1992
01:25:04,800 --> 01:25:09,920
So my job was to make the server pretty.

1993
01:25:07,199 --> 01:25:11,920
Alex over here wrote a basic client file

1994
01:25:09,920 --> 01:25:17,120
and I took it and made it pretty. So

1995
01:25:11,920 --> 01:25:19,199
it's now mine and legally distinct. So

1996
01:25:17,120 --> 01:25:22,239
the way I built it is using a library

1997
01:25:19,199 --> 01:25:26,000
called dear pi gy which is based off a

1998
01:25:22,239 --> 01:25:28,239
C++ library called dear I'm gy. And as

1999
01:25:26,000 --> 01:25:30,639
you can see here there's two input boxes

2000
01:25:28,239 --> 01:25:32,560
to input a username and a password. And

2001
01:25:30,639 --> 01:25:35,280
there's two buttons down here login and

2002
01:25:32,560 --> 01:25:37,360
register. So what login does is it will

2003
01:25:35,280 --> 01:25:40,080
take the username and password. It will

2004
01:25:37,360 --> 01:25:42,400
send it to the server and then it will

2005
01:25:40,080 --> 01:25:44,560
return a result whether it's it was

2006
01:25:42,400 --> 01:25:46,320
found in the database or not. What

2007
01:25:44,560 --> 01:25:48,239
register does is it takes the username

2008
01:25:46,320 --> 01:25:51,360
and password and makes a new login which

2009
01:25:48,239 --> 01:25:54,639
it then logs into. Once you log in or

2010
01:25:51,360 --> 01:25:56,400
register it takes you to a chat page. As

2011
01:25:54,639 --> 01:25:58,639
you can see here, there's a chat window

2012
01:25:56,400 --> 01:26:01,360
where you can see text from both you and

2013
01:25:58,639 --> 01:26:03,920
Siggment. And there's an input box where

2014
01:26:01,360 --> 01:26:06,560
you can type a message to Sigment or you

2015
01:26:03,920 --> 01:26:09,040
can hold that voice button to speak to

2016
01:26:06,560 --> 01:26:10,960
Sigment. But it doesn't, this library

2017
01:26:09,040 --> 01:26:13,920
doesn't really work on Dear Pi Gooey.

2018
01:26:10,960 --> 01:26:17,679
So, we've got like a

2019
01:26:13,920 --> 01:26:20,840
textbased GUI system. Over to Haley for

2020
01:26:17,679 --> 01:26:20,840
the face.

2021
01:26:21,840 --> 01:26:26,159
The reason we gave Sigman to face was to

2022
01:26:24,080 --> 01:26:27,520
differentiate ourselves from other AI

2023
01:26:26,159 --> 01:26:31,320
therapists that are already on the

2024
01:26:27,520 --> 01:26:31,320
market like

2025
01:26:31,840 --> 01:26:36,719
an actual example is on kids helpline

2026
01:26:34,639 --> 01:26:38,400
where they have an AI chatbot. We found

2027
01:26:36,719 --> 01:26:40,800
that these products while being useful

2028
01:26:38,400 --> 01:26:42,400
we thought that a face would allow um

2029
01:26:40,800 --> 01:26:43,920
the client to build a deeper connection

2030
01:26:42,400 --> 01:26:46,320
with it and improve outcomes for

2031
01:26:43,920 --> 01:26:48,239
therapy. So to further this, we tried to

2032
01:26:46,320 --> 01:26:50,159
make the face as emotional as possible

2033
01:26:48,239 --> 01:26:52,719
using things like eyebrows as well as

2034
01:26:50,159 --> 01:26:54,239
different colored LEDs. Uh we had the

2035
01:26:52,719 --> 01:26:56,719
best of intentions with this, but it

2036
01:26:54,239 --> 01:26:58,239
ended up just making it creepy as hell.

2037
01:26:56,719 --> 01:27:01,280
So

2038
01:26:58,239 --> 01:27:03,840
good job, I suppose.

2039
01:27:01,280 --> 01:27:06,000
So the way the eyebrows and eyelids work

2040
01:27:03,840 --> 01:27:08,000
is using I squared C to act as like a

2041
01:27:06,000 --> 01:27:10,080
middleman between the Raspberry Pi and

2042
01:27:08,000 --> 01:27:12,320
the servos. So we can just input the

2043
01:27:10,080 --> 01:27:14,639
angle we want them to be at rather than

2044
01:27:12,320 --> 01:27:18,400
have to put in the PW and signals

2045
01:27:14,639 --> 01:27:20,239
ourselves. And then with the RGB LEDs,

2046
01:27:18,400 --> 01:27:22,320
they change color based on the different

2047
01:27:20,239 --> 01:27:24,159
emotion of what like sign text is

2048
01:27:22,320 --> 01:27:27,679
showing. So we actually get the llama

2049
01:27:24,159 --> 01:27:29,760
model to perform uh textural analysis on

2050
01:27:27,679 --> 01:27:32,480
itself. Uh sentiment analysis, I forgot

2051
01:27:29,760 --> 01:27:34,400
the word. Um so that we can get

2052
01:27:32,480 --> 01:27:36,800
different emotions. So something like

2053
01:27:34,400 --> 01:27:38,800
blue for sad or yellow for happy. Moving

2054
01:27:36,800 --> 01:27:40,159
over to Ryland.

2055
01:27:38,800 --> 01:27:41,440
Great. So, um, if you actually want to

2056
01:27:40,159 --> 01:27:43,199
talk to Simmon, you have to have voice

2057
01:27:41,440 --> 01:27:44,960
recognition. Um, for us, we're just

2058
01:27:43,199 --> 01:27:47,120
using two main parts. We got, uh,

2059
01:27:44,960 --> 01:27:48,719
recording and transcribing. Um, the

2060
01:27:47,120 --> 01:27:50,880
recording parts of it is just using two

2061
01:27:48,719 --> 01:27:52,480
libraries, sound device and sound file,

2062
01:27:50,880 --> 01:27:54,880
where you can play and record your audio

2063
01:27:52,480 --> 01:27:55,920
or read and write your audio. Um, so we

2064
01:27:54,880 --> 01:27:57,600
use that microphone over there for

2065
01:27:55,920 --> 01:27:58,960
input, which has a lot of samples, and

2066
01:27:57,600 --> 01:28:01,040
we just shove all of those into a web

2067
01:27:58,960 --> 01:28:03,280
file. Um, so that's one part and then we

2068
01:28:01,040 --> 01:28:06,880
get to transcribing which we use farts

2069
01:28:03,280 --> 01:28:08,080
to whisper which is just a a local text

2070
01:28:06,880 --> 01:28:09,920
to speech module which will go through

2071
01:28:08,080 --> 01:28:11,760
that audio file and transcribe it into a

2072
01:28:09,920 --> 01:28:13,280
single string of text. That text would

2073
01:28:11,760 --> 01:28:15,120
go to the server and we generate a

2074
01:28:13,280 --> 01:28:17,600
response which is then we when we get to

2075
01:28:15,120 --> 01:28:19,120
the voice output. The voice output is

2076
01:28:17,600 --> 01:28:21,679
again just using two libraries. It's

2077
01:28:19,120 --> 01:28:23,920
very simple. Um goes through Piper which

2078
01:28:21,679 --> 01:28:25,600
has about 30 English voices and about 42

2079
01:28:23,920 --> 01:28:27,600
different languages we can choose from.

2080
01:28:25,600 --> 01:28:30,639
However, we only speak English so we

2081
01:28:27,600 --> 01:28:33,280
only focus on English. Um, so we give

2082
01:28:30,639 --> 01:28:35,120
that we give that to Piper and then it

2083
01:28:33,280 --> 01:28:36,800
that it turns that string of text into

2084
01:28:35,120 --> 01:28:39,920
another web file which we play through

2085
01:28:36,800 --> 01:28:42,320
wave and which is what you heard before.

2086
01:28:39,920 --> 01:28:45,800
So if we go to a more elaborate demo of

2087
01:28:42,320 --> 01:28:45,800
what we had before

2088
01:28:58,800 --> 01:29:02,639
I'm feeling very nervous about a

2089
01:29:00,639 --> 01:29:06,520
presentation I'm currently doing. What

2090
01:29:02,639 --> 01:29:06,520
can I do to feel better?

2091
01:29:19,360 --> 01:29:23,600
Pineipa, I can sense that you're putting

2092
01:29:21,440 --> 01:29:25,760
a lot of pressure on yourself for this

2093
01:29:23,600 --> 01:29:28,000
presentation, and it's completely

2094
01:29:25,760 --> 01:29:30,320
understandable to feel nervous when

2095
01:29:28,000 --> 01:29:32,719
facing a challenging situation like

2096
01:29:30,320 --> 01:29:35,280
this. Can you tell me more about what

2097
01:29:32,719 --> 01:29:37,040
specifically is causing your anxiety? Is

2098
01:29:35,280 --> 01:29:41,560
there something in particular that's

2099
01:29:37,040 --> 01:29:41,560
making you worry or feel uncertain?

2100
01:29:42,800 --> 01:29:46,440
Keep talking to him.

2101
01:29:47,760 --> 01:29:51,280
Speak your mind, Lonnie.

2102
01:29:49,520 --> 01:29:52,960
There's a lot of people staring at me

2103
01:29:51,280 --> 01:29:54,800
right now, and it's making me feel a

2104
01:29:52,960 --> 01:29:58,040
little self-conscious. What can I do to

2105
01:29:54,800 --> 01:29:58,040
feel better?

2106
01:30:06,800 --> 01:30:10,880
But it sounds like the attention from

2107
01:30:09,040 --> 01:30:13,199
others is contributing to your

2108
01:30:10,880 --> 01:30:16,000
nervousness. But also acknowledge that

2109
01:30:13,199 --> 01:30:18,800
as public speaking goes on, sometimes we

2110
01:30:16,000 --> 01:30:20,960
get caught up in our own reactions. And

2111
01:30:18,800 --> 01:30:22,880
this feeling of being self-conscious

2112
01:30:20,960 --> 01:30:25,040
might be a bit more about how you're

2113
01:30:22,880 --> 01:30:28,239
perceiving yourself in the moment rather

2114
01:30:25,040 --> 01:30:32,080
than an objective reality.

2115
01:30:28,239 --> 01:30:32,080
We think you're doing great.

2116
01:30:32,800 --> 01:30:37,239
So, I think that's it. I think that's

2117
01:30:34,239 --> 01:30:37,239
it.

2118
01:30:47,060 --> 01:30:55,350
[Applause]

2119
01:30:56,000 --> 01:31:00,320
Awesome.

2120
01:30:57,679 --> 01:31:03,320
Can I have the biggest Round of

2121
01:31:00,320 --> 01:31:03,320
applause.

2122
01:31:04,639 --> 01:31:08,800
Can I have the biggest round of Yes. The

2123
01:31:07,280 --> 01:31:09,920
biggest round of applause for this. And

2124
01:31:08,800 --> 01:31:11,199
you can leave it there for a second if

2125
01:31:09,920 --> 01:31:12,239
you want or you can take it because I'm

2126
01:31:11,199 --> 01:31:14,639
actually going to get the biggest round

2127
01:31:12,239 --> 01:31:16,880
of applause for all of the students to

2128
01:31:14,639 --> 01:31:20,400
stand up. Yeah.

2129
01:31:16,880 --> 01:31:23,679
Come on students. Come on up.

2130
01:31:20,400 --> 01:31:25,280
Come on up on stage.

2131
01:31:23,679 --> 01:31:30,480
Let's see what this stage was weight

2132
01:31:25,280 --> 01:31:32,880
rated for. um because

2133
01:31:30,480 --> 01:31:34,159
they all did a phenomenal job. And I

2134
01:31:32,880 --> 01:31:35,199
know there are questions in here. I'm

2135
01:31:34,159 --> 01:31:39,400
going to get out of the way so you can

2136
01:31:35,199 --> 01:31:39,400
all take photos. Um

2137
01:31:41,840 --> 01:31:46,080
excellent. You get fancy stickers. Uh

2138
01:31:44,560 --> 01:31:49,199
and

2139
01:31:46,080 --> 01:31:51,920
do they have what questions do we have

2140
01:31:49,199 --> 01:31:57,840
in the room?

2141
01:31:51,920 --> 01:31:57,840
Yes. Thank you. Here we go.

2142
01:31:58,800 --> 01:32:03,199
Uh so my question is for Leticia. Um

2143
01:32:01,199 --> 01:32:06,719
your one of the classifiers was based

2144
01:32:03,199 --> 01:32:09,520
upon the spectroscope data. How visually

2145
01:32:06,719 --> 01:32:11,360
distinct are those spec spectroscopic

2146
01:32:09,520 --> 01:32:13,120
images? Like could a human tell them

2147
01:32:11,360 --> 01:32:15,120
apart or is this something where the

2148
01:32:13,120 --> 01:32:16,639
neural network is seeing very very

2149
01:32:15,120 --> 01:32:20,080
subtle differences and is and is

2150
01:32:16,639 --> 01:32:24,280
differentiating based on that?

2151
01:32:20,080 --> 01:32:24,280
Hang on, hang on, hang on, hang on.

2152
01:32:28,719 --> 01:32:34,560
Um, it's called a spectrogram and some

2153
01:32:32,080 --> 01:32:37,520
songs you can tell them apart but others

2154
01:32:34,560 --> 01:32:40,520
it can capture like the nuances.

2155
01:32:37,520 --> 01:32:40,520
Yes,

2156
01:32:51,920 --> 01:32:56,880
this question is to the group that just

2157
01:32:54,560 --> 01:33:00,080
presented. Then

2158
01:32:56,880 --> 01:33:03,600
that that face is that based on any one

2159
01:33:00,080 --> 01:33:05,679
of four of you?

2160
01:33:03,600 --> 01:33:07,520
Uh, no. Actually, I got the face model

2161
01:33:05,679 --> 01:33:10,080
from Thingiverse. We did have an older

2162
01:33:07,520 --> 01:33:13,520
one that was based on him, but it didn't

2163
01:33:10,080 --> 01:33:17,239
end up working out. So, no,

2164
01:33:13,520 --> 01:33:17,239
it wasn't very pretty.

2165
01:33:26,400 --> 01:33:31,040
Uh this one's for Mars I think it was.

2166
01:33:28,320 --> 01:33:33,360
Um who spent your life uh doing back

2167
01:33:31,040 --> 01:33:36,719
propagation. Um did you consider using

2168
01:33:33,360 --> 01:33:38,239
anything like or number or jacks to try

2169
01:33:36,719 --> 01:33:40,159
and speed it up?

2170
01:33:38,239 --> 01:33:42,080
I have never heard of those things.

2171
01:33:40,159 --> 01:33:43,520
Okay.

2172
01:33:42,080 --> 01:33:44,400
I'm sorry to say it might have saved you

2173
01:33:43,520 --> 01:33:49,400
some time.

2174
01:33:44,400 --> 01:33:49,400
Um but I will look into it though. Yes.

2175
01:33:50,719 --> 01:33:57,840
Um this one's for Shri. How

2176
01:33:54,719 --> 01:34:01,120
how well did your uh model and analysis

2177
01:33:57,840 --> 01:34:04,560
compare to other models analyzing and

2178
01:34:01,120 --> 01:34:06,400
and classifying tumors

2179
01:34:04,560 --> 01:34:09,360
like other existing models?

2180
01:34:06,400 --> 01:34:11,280
Yeah. Like I'm I'm not I'm not expecting

2181
01:34:09,360 --> 01:34:13,679
you to be you know cutting edge although

2182
01:34:11,280 --> 01:34:15,840
if you are that's awesome but like you

2183
01:34:13,679 --> 01:34:19,280
you were getting sort of 90 96ish

2184
01:34:15,840 --> 01:34:21,120
percent I think. What are the other

2185
01:34:19,280 --> 01:34:23,920
models that you've heard of getting? Do

2186
01:34:21,120 --> 01:34:25,920
you know? So this was the first time I

2187
01:34:23,920 --> 01:34:27,520
did image classification. All the

2188
01:34:25,920 --> 01:34:30,239
projects before that I was just

2189
01:34:27,520 --> 01:34:33,600
classifying straight up CSVs which

2190
01:34:30,239 --> 01:34:36,080
weren't image data. And for those I've

2191
01:34:33,600 --> 01:34:38,719
experimented with using a bunch of um

2192
01:34:36,080 --> 01:34:43,040
libraries from sklearn

2193
01:34:38,719 --> 01:34:45,600
um like decision trees and linear

2194
01:34:43,040 --> 01:34:48,320
regression and they all had about the

2195
01:34:45,600 --> 01:34:49,920
same amount 96%. Um, for image

2196
01:34:48,320 --> 01:34:51,679
classification, I haven't actually

2197
01:34:49,920 --> 01:34:54,320
experimented with anything else, so I

2198
01:34:51,679 --> 01:34:57,840
might check that out now. But I haven't

2199
01:34:54,320 --> 01:35:02,080
really seen anything that scores above

2200
01:34:57,840 --> 01:35:06,040
96% to be honest. So, I think it it did

2201
01:35:02,080 --> 01:35:06,040
pretty good. Yeah,

2202
01:35:06,080 --> 01:35:10,639
I have a question on the Discord.

2203
01:35:08,080 --> 01:35:12,239
Yeah, that one. Uh, so for the link

2204
01:35:10,639 --> 01:35:14,480
talk, uh, you said you were going to

2205
01:35:12,239 --> 01:35:16,159
change, uh, so that the suggestions were

2206
01:35:14,480 --> 01:35:17,360
less esoteric. Do you have any thoughts

2207
01:35:16,159 --> 01:35:21,000
about

2208
01:35:17,360 --> 01:35:21,000
how to do that.

2209
01:35:21,679 --> 01:35:24,639
Oh, sorry. Okay,

2210
01:35:23,440 --> 01:35:26,560
that's okay.

2211
01:35:24,639 --> 01:35:29,360
Yeah,

2212
01:35:26,560 --> 01:35:31,920
if you Hudson's on the Discord if he can

2213
01:35:29,360 --> 01:35:33,920
answer that himself. Excellent. I like

2214
01:35:31,920 --> 01:35:36,000
this question that Anthony that I am

2215
01:35:33,920 --> 01:35:37,920
really bad at recogniz. But I I'm going

2216
01:35:36,000 --> 01:35:39,760
to put this question perhaps as it's

2217
01:35:37,920 --> 01:35:42,480
more homework than question. Anthony

2218
01:35:39,760 --> 01:35:44,800
asks, "What is your," and I assume

2219
01:35:42,480 --> 01:35:47,040
that's the plural, um, least favorite

2220
01:35:44,800 --> 01:35:50,430
thing about Python, and which of you

2221
01:35:47,040 --> 01:35:51,760
will fix it?

2222
01:35:50,430 --> 01:35:53,679
[Applause]

2223
01:35:51,760 --> 01:35:56,320
I also noticed that they're all standing

2224
01:35:53,679 --> 01:36:00,840
like right up against the screen, so no

2225
01:35:56,320 --> 01:36:00,840
one can step backwards first.

2226
01:36:00,880 --> 01:36:05,040
Uh, does anyone want to say what their

2227
01:36:03,440 --> 01:36:07,920
least thing about least favorite thing

2228
01:36:05,040 --> 01:36:10,800
about Python is? and how I think they

2229
01:36:07,920 --> 01:36:13,280
all love everything about Python.

2230
01:36:10,800 --> 01:36:16,400
Um, are there any other questions?

2231
01:36:13,280 --> 01:36:19,400
Otherwise, uh, yeah, two more questions.

2232
01:36:16,400 --> 01:36:19,400
Alyssa,

2233
01:36:20,560 --> 01:36:23,600
you have a Python core developer in the

2234
01:36:22,159 --> 01:36:27,719
room. So, is there anything you want to

2235
01:36:23,600 --> 01:36:27,719
know about contributing to CPython?

2236
01:36:31,040 --> 01:36:37,080
Does anyone else want to answer?

2237
01:36:34,080 --> 01:36:37,080
Could

2238
01:36:37,280 --> 01:36:40,960
you could you repeat the question,

2239
01:36:38,960 --> 01:36:42,800
please?

2240
01:36:40,960 --> 01:36:44,719
Again, I think the question was more

2241
01:36:42,800 --> 01:36:45,199
homework than a genuine question.

2242
01:36:44,719 --> 01:36:47,600
Okay.

2243
01:36:45,199 --> 01:36:50,719
Um, but if you'd like to learn how to

2244
01:36:47,600 --> 01:36:52,800
contribute to to Python itself,

2245
01:36:50,719 --> 01:36:56,400
essentially,

2246
01:36:52,800 --> 01:37:00,199
okay, come talk to these folks. Uh,

2247
01:36:56,400 --> 01:37:00,199
okay. Here we go.

2248
01:37:00,719 --> 01:37:07,920
Hey. Um, so this one's for Kevin.

2249
01:37:04,080 --> 01:37:09,840
Uh did you end up continuing with the uh

2250
01:37:07,920 --> 01:37:11,600
like Flask back end or did you end up

2251
01:37:09,840 --> 01:37:14,320
changing the the back end later on?

2252
01:37:11,600 --> 01:37:16,480
Oh yeah. Um so I ended up pivoting most

2253
01:37:14,320 --> 01:37:18,639
of the functionality to like a NetJS

2254
01:37:16,480 --> 01:37:22,719
back end but I'm still using Flask for

2255
01:37:18,639 --> 01:37:24,800
like my AI um MCP servers um which is um

2256
01:37:22,719 --> 01:37:27,119
AI I mean sorry Python has a larger

2257
01:37:24,800 --> 01:37:29,040
ecosystem for like machine learning as

2258
01:37:27,119 --> 01:37:31,440
well as AI. So I'm probably going to

2259
01:37:29,040 --> 01:37:34,880
stick with FL for those like um

2260
01:37:31,440 --> 01:37:34,880
functionalities. Yeah.

2261
01:37:40,480 --> 01:37:44,080
We're just conferring on the on more

2262
01:37:42,159 --> 01:37:45,679
questions

2263
01:37:44,080 --> 01:37:48,679
even though we've absolutely run out of

2264
01:37:45,679 --> 01:37:48,679
time.

2265
01:37:48,960 --> 01:37:52,320
Ah, and I've forgotten how to use the

2266
01:37:50,639 --> 01:37:54,080
internet.

2267
01:37:52,320 --> 01:37:54,400
Um,

2268
01:37:54,080 --> 01:37:57,440
yes.

2269
01:37:54,400 --> 01:38:02,320
So, for Kayn, uh, we just have a

2270
01:37:57,440 --> 01:38:05,119
question about the, um, CSS styling. Uh,

2271
01:38:02,320 --> 01:38:07,760
that's like pro.

2272
01:38:05,119 --> 01:38:08,960
So, people were asking how a a year 12

2273
01:38:07,760 --> 01:38:10,880
student gets to that level.

2274
01:38:08,960 --> 01:38:12,960
Oh, it's just Tailwind. Like, I don't

2275
01:38:10,880 --> 01:38:15,119
know. I didn't mention it, but like if

2276
01:38:12,960 --> 01:38:16,960
you just um like Yeah, Tailwind like

2277
01:38:15,119 --> 01:38:19,600
they support like CSS is like in line,

2278
01:38:16,960 --> 01:38:21,440
so you can it's really easy. Um, as long

2279
01:38:19,600 --> 01:38:22,800
like you just get the hang of it. Um,

2280
01:38:21,440 --> 01:38:23,920
yeah, just Tailwind. I don't know if you

2281
01:38:22,800 --> 01:38:26,080
want to like look it up, but it's a

2282
01:38:23,920 --> 01:38:29,080
library for CSS and yeah, it's really

2283
01:38:26,080 --> 01:38:29,080
good.

2284
01:38:29,600 --> 01:38:33,280
Okay, I've just heard the worst

2285
01:38:30,960 --> 01:38:35,040
question.

2286
01:38:33,280 --> 01:38:41,960
Um, it's for the last scene that

2287
01:38:35,040 --> 01:38:41,960
presented. Can you make him angry? Like,

2288
01:38:42,719 --> 01:38:46,000
now you see why it was the worst

2289
01:38:44,480 --> 01:38:48,080
question.

2290
01:38:46,000 --> 01:38:51,719
Look away if you don't want Sigman to

2291
01:38:48,080 --> 01:38:51,719
haunt your dreams.

2292
01:38:52,960 --> 01:38:58,320
Why do his cheeks glow? Like what is

2293
01:38:56,400 --> 01:39:01,920
Were you modeling this on a were you you

2294
01:38:58,320 --> 01:39:04,000
were trying to Why?

2295
01:39:01,920 --> 01:39:06,960
Uh the cheeks glow cuz that's like where

2296
01:39:04,000 --> 01:39:08,480
human blush is and stuff, but it's like

2297
01:39:06,960 --> 01:39:10,800
where else were we going to put like the

2298
01:39:08,480 --> 01:39:15,119
color for emotion?

2299
01:39:10,800 --> 01:39:17,600
Ah yes. Traditionally blushing blue.

2300
01:39:15,119 --> 01:39:19,360
Um I don't think we can make Sigman

2301
01:39:17,600 --> 01:39:22,920
angry, but we can definitely make him We

2302
01:39:19,360 --> 01:39:22,920
can. Oh god.

2303
01:39:23,199 --> 01:39:29,600
Wait, why do we have

2304
01:39:24,639 --> 01:39:31,920
I I think I think he's doing it live.

2305
01:39:29,600 --> 01:39:34,920
Like I think he's just programming it in

2306
01:39:31,920 --> 01:39:34,920
now.

2307
01:39:35,600 --> 01:39:41,760
I I I also have a question um for you. I

2308
01:39:40,400 --> 01:39:43,360
should have asked this question a lot

2309
01:39:41,760 --> 01:39:46,800
earlier. Have you sourced any of your

2310
01:39:43,360 --> 01:39:48,719
robot components from the tip?

2311
01:39:46,800 --> 01:39:50,800
Again, this is mostly a question for the

2312
01:39:48,719 --> 01:39:53,199
teacher in the room

2313
01:39:50,800 --> 01:39:56,159
and or a graveyard. Yeah, that's you

2314
01:39:53,199 --> 01:39:58,560
know, honestly, maybe. Yeah.

2315
01:39:56,159 --> 01:40:01,360
If you define parts of the classroom,

2316
01:39:58,560 --> 01:40:02,880
the tip. Yes.

2317
01:40:01,360 --> 01:40:06,080
Rude.

2318
01:40:02,880 --> 01:40:09,760
He's your teacher's right there.

2319
01:40:06,080 --> 01:40:13,520
My gosh. Um, okay. I look, I think we've

2320
01:40:09,760 --> 01:40:15,679
got we got we've got a few slides to say

2321
01:40:13,520 --> 01:40:17,679
thank you to everyone for coming. Uh,

2322
01:40:15,679 --> 01:40:21,080
oh, there's another qu I wasn't ignoring

2323
01:40:17,679 --> 01:40:21,080
you. You're just

2324
01:40:21,920 --> 01:40:25,760
Uh, this is a question for Evelyn. Just

2325
01:40:23,760 --> 01:40:27,520
on your musical graphing, I was

2326
01:40:25,760 --> 01:40:29,040
wondering if you had any other like

2327
01:40:27,520 --> 01:40:30,320
ambitions cuz like it's pretty cool to

2328
01:40:29,040 --> 01:40:31,920
see the bouncing, but like do you

2329
01:40:30,320 --> 01:40:35,119
imagine adding any other colors or other

2330
01:40:31,920 --> 01:40:37,360
like other modes of creation other than

2331
01:40:35,119 --> 01:40:39,760
just like the sine wave kind of shape?

2332
01:40:37,360 --> 01:40:43,360
Maybe some emoji on the Does Does this

2333
01:40:39,760 --> 01:40:45,760
work? Does this work? Oh, yeah. So, Huh?

2334
01:40:43,360 --> 01:40:49,119
Oh, I kind of mentioned it uh like

2335
01:40:45,760 --> 01:40:52,480
really briefly at the end, but um yeah,

2336
01:40:49,119 --> 01:40:55,440
since the sine wave can only like show

2337
01:40:52,480 --> 01:40:57,520
the five frequency groups, I do want to

2338
01:40:55,440 --> 01:40:59,199
change it to a different visualizer so I

2339
01:40:57,520 --> 01:41:02,639
can show the different frequency groups.

2340
01:40:59,199 --> 01:41:04,880
Colors would be nice. Um

2341
01:41:02,639 --> 01:41:08,440
yeah, so I do plan to do that. That

2342
01:41:04,880 --> 01:41:08,440
would be good.

2343
01:41:11,280 --> 01:41:16,360
Okay. Um

2344
01:41:13,360 --> 01:41:16,360
Applause.