1
00:00:02,750 --> 00:00:08,520
everyone so this other sort of wave a

2
00:00:06,029 --> 00:00:12,330
question and that's the question what

3
00:00:08,520 --> 00:00:14,480
does Minority Report Black Mirror a 1984

4
00:00:12,330 --> 00:00:16,860
all have in common

5
00:00:14,480 --> 00:00:18,600
now it's not the fact that they're all

6
00:00:16,860 --> 00:00:21,270
forms of media you know books films TV

7
00:00:18,600 --> 00:00:24,300
shows nor is it the fact that are about

8
00:00:21,270 --> 00:00:26,310
dystopian futures but in says the fact

9
00:00:24,300 --> 00:00:29,789
that they each talked about predicting

10
00:00:26,310 --> 00:00:32,009
crime in one form or another whether

11
00:00:29,789 --> 00:00:33,239
that's the precogs in minority reports

12
00:00:32,009 --> 00:00:35,910
the Ricola

13
00:00:33,239 --> 00:00:39,089
in black mirror or the for police in

14
00:00:35,910 --> 00:00:41,370
1984 each of these forms of media look

15
00:00:39,090 --> 00:00:44,100
at how we could predict crime but more

16
00:00:41,370 --> 00:00:46,140
specifically the repercussions of doing

17
00:00:44,100 --> 00:00:46,800
so and that's what I'm going to be

18
00:00:46,140 --> 00:00:48,809
talking about today

19
00:00:46,800 --> 00:00:52,470
and we talk about how we can use natural

20
00:00:48,809 --> 00:00:53,640
language processing to predict crime but

21
00:00:52,470 --> 00:00:56,489
for those of you that know me well know

22
00:00:53,640 --> 00:00:59,070
I'm not a mathematician and also not a

23
00:00:56,489 --> 00:01:00,870
police officer so why am I talking about

24
00:00:59,070 --> 00:01:03,120
natural language processing which is

25
00:01:00,870 --> 00:01:05,548
quite mathy and predictive policing

26
00:01:03,120 --> 00:01:08,249
which has the name may suggest it's all

27
00:01:05,549 --> 00:01:10,530
to do with law with crime what it comes

28
00:01:08,249 --> 00:01:13,530
down to this quote the idea that

29
00:01:10,530 --> 00:01:15,929
intrusion analysis security analysis

30
00:01:13,530 --> 00:01:19,049
it's about far more than the tools we

31
00:01:15,929 --> 00:01:20,579
use it's about everything and looking at

32
00:01:19,049 --> 00:01:23,399
new ways that we can protect ourselves

33
00:01:20,579 --> 00:01:26,669
from attacks but also predict those

34
00:01:23,399 --> 00:01:27,869
attacks in the first place so what am I

35
00:01:26,670 --> 00:01:29,609
actually going to be talking about today

36
00:01:27,869 --> 00:01:31,319
well I want to break it up into three

37
00:01:29,609 --> 00:01:33,990
main areas I don't talk about what

38
00:01:31,319 --> 00:01:35,069
predictive policing actually is I want

39
00:01:33,990 --> 00:01:37,439
to talk about what natural language

40
00:01:35,069 --> 00:01:38,670
processing is and then finally I wanna

41
00:01:37,439 --> 00:01:41,369
talk about how we can merge these two

42
00:01:38,670 --> 00:01:44,119
ideas together how we can predict crime

43
00:01:41,369 --> 00:01:46,109
using natural language processing

44
00:01:44,119 --> 00:01:48,060
provide you know

45
00:01:46,109 --> 00:01:50,039
my name is James Stevenson and this time

46
00:01:48,060 --> 00:01:51,840
two years ago now I was a student at the

47
00:01:50,039 --> 00:01:54,539
University of South Wales spitting

48
00:01:51,840 --> 00:01:56,429
convenience security before that I was

49
00:01:54,539 --> 00:01:58,349
an intern at alert logic a cloud

50
00:01:56,429 --> 00:02:00,859
security company and these days I'm a

51
00:01:58,349 --> 00:02:03,299
software engineer at BT security

52
00:02:00,859 --> 00:02:05,609
Praetorians ready alert what is

53
00:02:03,299 --> 00:02:07,160
predictive policing I keep thinking

54
00:02:05,609 --> 00:02:08,970
about it what would actually isn't

55
00:02:07,160 --> 00:02:11,640
because if we're going to use natural

56
00:02:08,970 --> 00:02:13,230
language processing to predict crime but

57
00:02:11,640 --> 00:02:15,940
we kind of need to know what predictive

58
00:02:13,230 --> 00:02:17,579
policing it's

59
00:02:15,940 --> 00:02:19,950
and it comes down to two main areas

60
00:02:17,580 --> 00:02:23,560
location-based predictive policing and

61
00:02:19,950 --> 00:02:25,149
individual based predictiveness now

62
00:02:23,560 --> 00:02:26,710
location-based predictive policing as

63
00:02:25,150 --> 00:02:29,890
the name suggests is all about looking

64
00:02:26,710 --> 00:02:33,730
at an area so saying in this area in the

65
00:02:29,890 --> 00:02:35,109
future is a crime likely took it now

66
00:02:33,730 --> 00:02:37,510
this map is a great example of

67
00:02:35,110 --> 00:02:39,130
location-based predictive policing so

68
00:02:37,510 --> 00:02:41,200
this is a map of London between a

69
00:02:39,130 --> 00:02:43,359
specific time period or the darker the

70
00:02:41,200 --> 00:02:45,790
color in terms that the more crime has a

71
00:02:43,360 --> 00:02:48,600
kid now this is a great example because

72
00:02:45,790 --> 00:02:51,579
we can say okay if a crime has occurred

73
00:02:48,600 --> 00:02:53,680
under these specific circumstances in

74
00:02:51,580 --> 00:02:55,570
the past but it means that crime isn't

75
00:02:53,680 --> 00:02:57,610
likely to occur under these same

76
00:02:55,570 --> 00:02:59,440
circumstances in the future

77
00:02:57,610 --> 00:03:01,720
now today were actually gonna be

78
00:02:59,440 --> 00:03:04,120
focusing on individual-based predictive

79
00:03:01,720 --> 00:03:06,250
policing now individual based predictive

80
00:03:04,120 --> 00:03:09,430
policing is all about looking at an

81
00:03:06,250 --> 00:03:12,520
individual and saying how likely is this

82
00:03:09,430 --> 00:03:14,020
individual to Connecticut and when it

83
00:03:12,520 --> 00:03:15,970
comes to this type of predictive

84
00:03:14,020 --> 00:03:18,520
policing with masks different questions

85
00:03:15,970 --> 00:03:21,630
will go down different routes different

86
00:03:18,520 --> 00:03:24,070
avenues I will be left with that school

87
00:03:21,630 --> 00:03:26,590
no today we're going to focus on three

88
00:03:24,070 --> 00:03:29,260
types or three approaches that can allow

89
00:03:26,590 --> 00:03:30,880
us to do justice the first of these

90
00:03:29,260 --> 00:03:33,970
approaches we're gonna look at is called

91
00:03:30,880 --> 00:03:37,000
strain theory now strain theory is the

92
00:03:33,970 --> 00:03:39,700
idea that society puts pressure on

93
00:03:37,000 --> 00:03:42,670
individuals to achieve specific thoughts

94
00:03:39,700 --> 00:03:44,679
like the American dream but when

95
00:03:42,670 --> 00:03:48,100
individuals lack the means to achieve

96
00:03:44,680 --> 00:03:51,250
those goals the more likely técnicos so

97
00:03:48,100 --> 00:03:52,480
that they can achieve them the next

98
00:03:51,250 --> 00:03:55,480
theory that we're gonna look at is

99
00:03:52,480 --> 00:03:58,329
called social control here now social

100
00:03:55,480 --> 00:04:00,730
control theory is the idea that people

101
00:03:58,330 --> 00:04:02,890
who lack close relationships commitments

102
00:04:00,730 --> 00:04:04,690
values on roles are more likely to

103
00:04:02,890 --> 00:04:07,510
commit crimes because they don't have

104
00:04:04,690 --> 00:04:10,989
those relationships or values as an

105
00:04:07,510 --> 00:04:12,489
anchor in society and the final theory

106
00:04:10,989 --> 00:04:15,130
that we're going to look at today it's

107
00:04:12,489 --> 00:04:17,168
called social disorganization view now

108
00:04:15,130 --> 00:04:20,798
what this very states is it says that

109
00:04:17,168 --> 00:04:22,990
location is key if you live for a work

110
00:04:20,798 --> 00:04:25,929
in an area known for a specific type of

111
00:04:22,990 --> 00:04:28,270
crime this theory states an intrinsic

112
00:04:25,930 --> 00:04:30,159
link by being that the more likely

113
00:04:28,270 --> 00:04:31,748
definitely

114
00:04:30,159 --> 00:04:33,909
so so far we've looked at what

115
00:04:31,749 --> 00:04:36,069
predictive policing is different types

116
00:04:33,909 --> 00:04:38,199
of predictive policing and how we can

117
00:04:36,069 --> 00:04:40,869
use predictive policing approaches to

118
00:04:38,199 --> 00:04:42,879
predict call but this talk is all about

119
00:04:40,869 --> 00:04:45,129
natural language process so about how we

120
00:04:42,879 --> 00:04:48,069
can use natural language processing to

121
00:04:45,129 --> 00:04:50,050
do just that but before we talk about

122
00:04:48,069 --> 00:04:53,619
natural language processing we need to

123
00:04:50,050 --> 00:04:55,899
talk about language this serve us as

124
00:04:53,619 --> 00:04:58,239
human beings language comes down to

125
00:04:55,899 --> 00:05:01,179
these three main theories speaking

126
00:04:58,239 --> 00:05:04,149
reading and writing things that we all

127
00:05:01,179 --> 00:05:06,099
do every day so because we do these

128
00:05:04,149 --> 00:05:08,259
things every day most of us or maybe

129
00:05:06,099 --> 00:05:12,459
some of us will be able to answer this

130
00:05:08,259 --> 00:05:19,139
question Harris - grunts plus England

131
00:05:12,459 --> 00:05:22,749
equals what now the answer isn't London

132
00:05:19,139 --> 00:05:25,689
because her sister France has London is

133
00:05:22,749 --> 00:05:28,089
- ain't it now if we didn't know that

134
00:05:25,689 --> 00:05:30,909
that was the answer why did we know that

135
00:05:28,089 --> 00:05:32,289
was it well we would have known that

136
00:05:30,909 --> 00:05:35,199
that was the answer because of the

137
00:05:32,289 --> 00:05:38,378
experiences of head we've read books

138
00:05:35,199 --> 00:05:40,479
go on the internet spoken to people and

139
00:05:38,379 --> 00:05:43,689
that's all built on knowledge base and

140
00:05:40,479 --> 00:05:45,550
our understanding so then if we were to

141
00:05:43,689 --> 00:05:47,919
give that question to our natural

142
00:05:45,550 --> 00:05:51,219
language processing machine would it be

143
00:05:47,919 --> 00:05:54,219
able to answer it well yes but only if

144
00:05:51,219 --> 00:05:57,009
we gave it the right context so this is

145
00:05:54,219 --> 00:05:59,079
the Wikipedia article for London and if

146
00:05:57,009 --> 00:06:00,909
we fed this into our natural language

147
00:05:59,079 --> 00:06:03,669
processing machine you would learn from

148
00:06:00,909 --> 00:06:06,279
that surrounding context northern that

149
00:06:03,669 --> 00:06:09,448
London is a city the only in that London

150
00:06:06,279 --> 00:06:11,349
is in the UK of which England is as well

151
00:06:09,449 --> 00:06:15,369
building that knowledge base and

152
00:06:11,349 --> 00:06:16,959
building down so that's how natural

153
00:06:15,369 --> 00:06:19,569
language processing works had a

154
00:06:16,959 --> 00:06:21,069
sentiment analysis week because

155
00:06:19,569 --> 00:06:23,289
sentiment analysis is all about that

156
00:06:21,069 --> 00:06:25,719
tool looking at a specific piece of text

157
00:06:23,289 --> 00:06:29,318
and saying what is the emotion what is

158
00:06:25,719 --> 00:06:31,329
the sentiment behind that text however

159
00:06:29,319 --> 00:06:34,569
it comes to us as human beings we have

160
00:06:31,329 --> 00:06:36,519
eight main pillars to our emotions but

161
00:06:34,569 --> 00:06:37,300
the sentiment analysis really really

162
00:06:36,519 --> 00:06:40,319
care about

163
00:06:37,300 --> 00:06:43,360
- that's positive emotions and negative

164
00:06:40,319 --> 00:06:43,790
emotions so how do we translate those

165
00:06:43,360 --> 00:06:47,450
eight

166
00:06:43,790 --> 00:06:49,520
hell is down to two well well we're

167
00:06:47,450 --> 00:06:51,590
talking about positive emotions we're

168
00:06:49,520 --> 00:06:55,099
really talking about trust joy anger

169
00:06:51,590 --> 00:06:56,210
this price and given the red herring and

170
00:06:55,100 --> 00:06:58,160
what we're talking about negative

171
00:06:56,210 --> 00:07:02,870
emotions we're talking about disgust

172
00:06:58,160 --> 00:07:04,790
sadness fear and anticipation and so if

173
00:07:02,870 --> 00:07:05,960
those are the emotions that we're

174
00:07:04,790 --> 00:07:07,760
talking about when we referred to

175
00:07:05,960 --> 00:07:11,960
sentiment analysis how do we actually

176
00:07:07,760 --> 00:07:13,370
get that emotion from text well if the

177
00:07:11,960 --> 00:07:16,460
same as in most machine learning

178
00:07:13,370 --> 00:07:17,990
approaches we take a massive dataset now

179
00:07:16,460 --> 00:07:20,659
that data set for us it's going to be

180
00:07:17,990 --> 00:07:23,240
fresh drawn abuse each element in that

181
00:07:20,660 --> 00:07:26,420
set is broken down into two sub sections

182
00:07:23,240 --> 00:07:29,570
the actual review and then the sentiment

183
00:07:26,420 --> 00:07:32,000
of that review for example I love my

184
00:07:29,570 --> 00:07:34,540
local pizza restaurant positive sentence

185
00:07:32,000 --> 00:07:37,670
well this place has really gone downhill

186
00:07:34,540 --> 00:07:40,640
negative sentence we then break this

187
00:07:37,670 --> 00:07:43,760
dataset down into we have our training

188
00:07:40,640 --> 00:07:45,409
set and we have our testing set when it

189
00:07:43,760 --> 00:07:47,450
comes to training our natural language

190
00:07:45,410 --> 00:07:49,130
processing machine we ask it to go

191
00:07:47,450 --> 00:07:51,110
through those entities and to have a

192
00:07:49,130 --> 00:07:53,180
look at what keywords are more prominent

193
00:07:51,110 --> 00:07:55,340
with a positive sentiment and what

194
00:07:53,180 --> 00:07:58,820
keywords are more prominent be negative

195
00:07:55,340 --> 00:07:59,929
certain then when it comes to testing we

196
00:07:58,820 --> 00:08:02,990
ask it to go through the remaining

197
00:07:59,930 --> 00:08:05,270
entities and fit for it to tell us what

198
00:08:02,990 --> 00:08:06,950
the sentiment is and then if that

199
00:08:05,270 --> 00:08:09,770
matches the sentiment we know them to

200
00:08:06,950 --> 00:08:13,190
have cooked so if it doesn't it mean

201
00:08:09,770 --> 00:08:15,109
something as well more so that's how

202
00:08:13,190 --> 00:08:17,060
natural language processing works

203
00:08:15,110 --> 00:08:19,340
that's have sentiment analysis works

204
00:08:17,060 --> 00:08:20,470
what are it exists what are some

205
00:08:19,340 --> 00:08:24,560
examples of natural language processing

206
00:08:20,470 --> 00:08:26,480
in the real well this is a WS comprehend

207
00:08:24,560 --> 00:08:29,360
or specifically comprehending medical

208
00:08:26,480 --> 00:08:31,010
which is a delicious approach to natural

209
00:08:29,360 --> 00:08:34,400
language processing when it comes to

210
00:08:31,010 --> 00:08:36,229
medicine and healthcare a doctor or

211
00:08:34,400 --> 00:08:38,449
health care professional will type in a

212
00:08:36,229 --> 00:08:40,550
patient's details symptoms information

213
00:08:38,450 --> 00:08:42,770
the natural language processing tool

214
00:08:40,549 --> 00:08:44,810
will go off do its thing and it will

215
00:08:42,770 --> 00:08:45,800
come back with T bits of information the

216
00:08:44,810 --> 00:08:49,189
things that have that health care

217
00:08:45,800 --> 00:08:51,770
professional Austin it next we have

218
00:08:49,190 --> 00:08:53,440
Taylor a I know Taylor I was Microsoft's

219
00:08:51,770 --> 00:08:57,050
approach to natural language processing

220
00:08:53,440 --> 00:08:59,180
when it came to a Twitter chat it

221
00:08:57,050 --> 00:09:01,790
Tarek's response to people depending on

222
00:08:59,180 --> 00:09:03,649
how people smoked it now it's quite

223
00:09:01,790 --> 00:09:06,370
controversial it lasted just under 24

224
00:09:03,649 --> 00:09:08,930
hours but nonetheless is a great example

225
00:09:06,370 --> 00:09:11,570
and then finally we have predictive text

226
00:09:08,930 --> 00:09:13,670
so overnight you had an Android or an

227
00:09:11,570 --> 00:09:15,560
iPhone where that predictive text this

228
00:09:13,670 --> 00:09:19,219
probably works in your device by using

229
00:09:15,560 --> 00:09:21,339
natural language process so there we

230
00:09:19,220 --> 00:09:24,680
have three great examples so healthcare

231
00:09:21,339 --> 00:09:26,450
communications and mobile phones but

232
00:09:24,680 --> 00:09:28,310
none of those examples look at how we

233
00:09:26,450 --> 00:09:30,410
can use natural language processing to

234
00:09:28,310 --> 00:09:33,050
predict decline which is why we're here

235
00:09:30,410 --> 00:09:36,079
for this talk so let's do that

236
00:09:33,050 --> 00:09:38,630
well this is Alice and it's Alice's job

237
00:09:36,079 --> 00:09:40,849
to do just that it's Alice's job to

238
00:09:38,630 --> 00:09:43,399
predict crime where she currently does

239
00:09:40,850 --> 00:09:45,019
this she individually and manually goes

240
00:09:43,399 --> 00:09:47,570
to different web sites chat forums

241
00:09:45,019 --> 00:09:49,610
social media accounts and two profiles

242
00:09:47,570 --> 00:09:52,550
individuals on their likelihood of

243
00:09:49,610 --> 00:09:55,490
commit crime but that's slow and

244
00:09:52,550 --> 00:09:58,189
laborious so how can we take this to the

245
00:09:55,490 --> 00:10:00,110
next level well we can automate it we

246
00:09:58,190 --> 00:10:02,510
can scrape these websites the same

247
00:10:00,110 --> 00:10:04,220
information that would help we can use

248
00:10:02,510 --> 00:10:06,620
natural language processing on the

249
00:10:04,220 --> 00:10:09,200
response and then we can return to Alice

250
00:10:06,620 --> 00:10:13,070
a school a score of how likely that

251
00:10:09,200 --> 00:10:14,209
individual is to commit a crime and this

252
00:10:13,070 --> 00:10:16,130
is what we're gonna be focusing on for

253
00:10:14,209 --> 00:10:17,329
the rest of the day we're going to be

254
00:10:16,130 --> 00:10:19,670
talking about how we could build a

255
00:10:17,329 --> 00:10:22,819
conceptual framework that allows us to

256
00:10:19,670 --> 00:10:25,000
do just this so if we were to create

257
00:10:22,820 --> 00:10:28,220
this framework what would we need to do

258
00:10:25,000 --> 00:10:31,160
well first of all Alice we need to sit

259
00:10:28,220 --> 00:10:33,470
down and decide on the impact for we'll

260
00:10:31,160 --> 00:10:36,170
be profiling and what is the impact of

261
00:10:33,470 --> 00:10:38,300
those individuals if that individual was

262
00:10:36,170 --> 00:10:40,810
to commit that crime or was to perform

263
00:10:38,300 --> 00:10:42,829
that attacked what would the impact be

264
00:10:40,810 --> 00:10:44,359
and this comes down for those three main

265
00:10:42,829 --> 00:10:48,890
areas the loss of confidentiality

266
00:10:44,360 --> 00:10:50,600
integrity and available once we have

267
00:10:48,890 --> 00:10:52,970
that impact we will work out our

268
00:10:50,600 --> 00:10:56,029
likelihood what is the likelihood of

269
00:10:52,970 --> 00:10:58,040
that individual committing that crime or

270
00:10:56,029 --> 00:10:59,120
performing that attack well this is

271
00:10:58,040 --> 00:11:00,980
where we're going to go back to those

272
00:10:59,120 --> 00:11:03,829
predictive policing approaches that we

273
00:11:00,980 --> 00:11:06,320
mentioned earlier on first of all you

274
00:11:03,829 --> 00:11:08,120
scrape these websites for that text we

275
00:11:06,320 --> 00:11:10,520
use natural language processing on the

276
00:11:08,120 --> 00:11:12,680
response and now we're saying

277
00:11:10,520 --> 00:11:15,949
does that text continue reference to any

278
00:11:12,680 --> 00:11:19,310
goals or aspirations and then if it does

279
00:11:15,950 --> 00:11:21,650
what is the sentiment next we take that

280
00:11:19,310 --> 00:11:24,020
same bit of text we see now does that

281
00:11:21,650 --> 00:11:26,660
text contain reference to any close

282
00:11:24,020 --> 00:11:30,350
relationships any individuals any groups

283
00:11:26,660 --> 00:11:33,170
in the organizations and if so what is

284
00:11:30,350 --> 00:11:35,510
the sentiment and then finally we take

285
00:11:33,170 --> 00:11:37,219
that same bit of text once again and we

286
00:11:35,510 --> 00:11:40,790
say now does that text contain reference

287
00:11:37,220 --> 00:11:42,860
to the individuals location if so is

288
00:11:40,790 --> 00:11:45,430
that the location no that type of crime

289
00:11:42,860 --> 00:11:47,680
and then finally what is the sentiment

290
00:11:45,430 --> 00:11:50,839
we think of through each of these trees

291
00:11:47,680 --> 00:11:54,589
aggregating a score as we go and that

292
00:11:50,840 --> 00:11:57,110
score is our overall like it we can then

293
00:11:54,590 --> 00:11:59,660
use that whoever our impact to get our

294
00:11:57,110 --> 00:12:03,830
risk and that risk is the risk that this

295
00:11:59,660 --> 00:12:05,630
individual poses to Alice entity now

296
00:12:03,830 --> 00:12:07,700
that we have that risk it's just about

297
00:12:05,630 --> 00:12:09,820
collecting too much information as we

298
00:12:07,700 --> 00:12:12,140
can using natural language processing

299
00:12:09,820 --> 00:12:15,400
moon collect information like come on

300
00:12:12,140 --> 00:12:17,960
topics trans age gender and race

301
00:12:15,400 --> 00:12:21,140
occupation salary and religion and then

302
00:12:17,960 --> 00:12:23,180
any dates and times now the reason why

303
00:12:21,140 --> 00:12:25,189
we haven't focused on this information

304
00:12:23,180 --> 00:12:27,979
today it's because this information has

305
00:12:25,190 --> 00:12:30,740
the scope and the potential of becoming

306
00:12:27,980 --> 00:12:35,540
significantly more bucks and that's

307
00:12:30,740 --> 00:12:37,250
really a talk for another day so finally

308
00:12:35,540 --> 00:12:39,199
as part of this conceptual framework

309
00:12:37,250 --> 00:12:41,750
it's about creating a naming convention

310
00:12:39,200 --> 00:12:44,000
an image invention that we can instantly

311
00:12:41,750 --> 00:12:45,860
clean information from without including

312
00:12:44,000 --> 00:12:49,280
any of that personally identifiable

313
00:12:45,860 --> 00:12:51,740
information so this name is broken down

314
00:12:49,280 --> 00:12:54,949
into four main areas the source of the

315
00:12:51,740 --> 00:12:57,170
data the time it occurred reverse score

316
00:12:54,950 --> 00:13:01,160
and then a pseudo-random word that gives

317
00:12:57,170 --> 00:13:03,260
the name some uniqueness so we looked at

318
00:13:01,160 --> 00:13:05,630
what predictive policing is we look at

319
00:13:03,260 --> 00:13:07,040
what natural language processing is we

320
00:13:05,630 --> 00:13:11,330
also look at how we can use these two

321
00:13:07,040 --> 00:13:12,500
ideas together to predict crime but you

322
00:13:11,330 --> 00:13:14,180
might be thinking well James that's

323
00:13:12,500 --> 00:13:16,190
great but why are we talking about

324
00:13:14,180 --> 00:13:18,800
predictive policing and a computer

325
00:13:16,190 --> 00:13:21,440
security conference and well it comes

326
00:13:18,800 --> 00:13:22,990
back to this quote the idea that

327
00:13:21,440 --> 00:13:25,270
intrusion analysis

328
00:13:22,990 --> 00:13:27,820
security analysis it's about far more

329
00:13:25,270 --> 00:13:29,740
than the tools we use it's about in

330
00:13:27,820 --> 00:13:32,410
evicting but it's also about thinking

331
00:13:29,740 --> 00:13:34,120
outside of the box and looking at new

332
00:13:32,410 --> 00:13:36,910
ways that we can protect ourselves from

333
00:13:34,120 --> 00:13:40,450
attacks but also predict those attacks

334
00:13:36,910 --> 00:13:41,980
in the first place so already knows I'm

335
00:13:40,450 --> 00:13:43,570
gonna go through some questions that I

336
00:13:41,980 --> 00:13:45,730
only get that's one of this book

337
00:13:43,570 --> 00:13:47,589
I'll close the talk off and then if we

338
00:13:45,730 --> 00:13:50,020
have any additional questions I'll give

339
00:13:47,589 --> 00:13:51,520
written them so the first question we

340
00:13:50,020 --> 00:13:54,640
have here is is predictive policing

341
00:13:51,520 --> 00:13:56,740
better than what is now I think the

342
00:13:54,640 --> 00:13:59,020
answer here is no predictive policing is

343
00:13:56,740 --> 00:14:00,970
a tool it's a supplement it's something

344
00:13:59,020 --> 00:14:02,649
that should be used in addition to

345
00:14:00,970 --> 00:14:04,810
normal placing and isn't there to

346
00:14:02,649 --> 00:14:07,360
replace police nor is it that to replace

347
00:14:04,810 --> 00:14:10,479
police intuition the next question we

348
00:14:07,360 --> 00:14:12,160
have is predictive policing bias again

349
00:14:10,480 --> 00:14:15,550
the short answer is yes

350
00:14:12,160 --> 00:14:17,319
but even policing is quite bias the

351
00:14:15,550 --> 00:14:21,310
prison for that is predictive policing

352
00:14:17,320 --> 00:14:24,760
is garbage in garbage out if our data is

353
00:14:21,310 --> 00:14:27,369
bias it means how frameworks are going

354
00:14:24,760 --> 00:14:29,290
to be biased also and the problem with

355
00:14:27,370 --> 00:14:31,899
crime data is that it's intrinsically

356
00:14:29,290 --> 00:14:34,329
bias because we have crimes that aren't

357
00:14:31,899 --> 00:14:36,459
documented maybe like assault means that

358
00:14:34,329 --> 00:14:39,040
the data we get from that isn't accurate

359
00:14:36,459 --> 00:14:42,670
is it representative of the real world

360
00:14:39,040 --> 00:14:45,250
and that's fine as long as we remember

361
00:14:42,670 --> 00:14:46,479
the answer to that first question as

362
00:14:45,250 --> 00:14:49,510
long as we remember that predictive

363
00:14:46,480 --> 00:14:51,670
policing is a tool it's a suite and it's

364
00:14:49,510 --> 00:14:54,790
to be used in addition to what we're

365
00:14:51,670 --> 00:14:56,349
here to do next question there is is

366
00:14:54,790 --> 00:14:58,390
predictive policing used in the real

367
00:14:56,350 --> 00:15:00,430
world so pretty deep listens used in the

368
00:14:58,390 --> 00:15:01,660
UK it's also used in the u.s. one

369
00:15:00,430 --> 00:15:03,790
example that received quite a lot of

370
00:15:01,660 --> 00:15:04,270
media attention was the LAPD in the

371
00:15:03,790 --> 00:15:06,790
States

372
00:15:04,270 --> 00:15:09,010
they had a scheme called laser the way

373
00:15:06,790 --> 00:15:09,760
that laser which is assigned a score to

374
00:15:09,010 --> 00:15:11,430
ex-offenders

375
00:15:09,760 --> 00:15:14,380
that school was based on their behavior

376
00:15:11,430 --> 00:15:16,839
and the way that that worked is anyone

377
00:15:14,380 --> 00:15:20,260
in that top bracket would then receive a

378
00:15:16,839 --> 00:15:22,390
follow-up visit crumples the penultimate

379
00:15:20,260 --> 00:15:24,430
question we have here is how good is

380
00:15:22,390 --> 00:15:27,339
natural language processing at picking

381
00:15:24,430 --> 00:15:29,079
up differences or nuances in text and

382
00:15:27,339 --> 00:15:31,000
there's an example I like living for

383
00:15:29,079 --> 00:15:32,920
this and it was an example of a natural

384
00:15:31,000 --> 00:15:35,290
language processing tool that was

385
00:15:32,920 --> 00:15:36,490
created to simultaneously understand two

386
00:15:35,290 --> 00:15:37,630
different languages

387
00:15:36,490 --> 00:15:39,580
I think that's really interesting

388
00:15:37,630 --> 00:15:42,100
because that goes to share than actually

389
00:15:39,580 --> 00:15:44,020
in some cases natural language

390
00:15:42,100 --> 00:15:46,210
processing can be better at

391
00:15:44,020 --> 00:15:48,970
understanding text than some of us as

392
00:15:46,210 --> 00:15:50,980
human beings and finally what's next

393
00:15:48,970 --> 00:15:52,660
well specifically for this framework is

394
00:15:50,980 --> 00:15:54,460
about creating a proof of concept a

395
00:15:52,660 --> 00:15:56,560
prove a concept that allows us to do

396
00:15:54,460 --> 00:15:58,750
just this allows us to look at risk

397
00:15:56,560 --> 00:16:01,270
levels for individuals that allows us to

398
00:15:58,750 --> 00:16:03,370
pivot off social media information and

399
00:16:01,270 --> 00:16:05,140
allows it to a power ottoman and allows

400
00:16:03,370 --> 00:16:08,860
us to look at crime signals so

401
00:16:05,140 --> 00:16:10,750
indicators the cry however that is this

402
00:16:08,860 --> 00:16:12,640
talk came to a close if you do have any

403
00:16:10,750 --> 00:16:14,020
questions feel free to ask me now come

404
00:16:12,640 --> 00:16:15,880
find me afterwards or i'm also on

405
00:16:14,020 --> 00:16:18,840
twitter and underscore james deals

406
00:16:15,880 --> 00:16:18,840
thanks