WEBVTT

1
00:00:01.199 --> 00:00:06.200
<v Speaker 1>Welcome to the sentient Code, where intelligence is engineered, autonomy

2
00:00:06.280 --> 00:00:10.439
<v Speaker 1>is emerging, and a line between human and machine grows thinner.

3
00:00:10.800 --> 00:00:15.359
<v Speaker 1>Each episode, we decode the algorithms, explore the robotics, and

4
00:00:15.439 --> 00:00:20.519
<v Speaker 1>examine the ideas shaping the future of artificial minds.

5
00:00:23.920 --> 00:00:27.719
<v Speaker 2>Imagine you've just found your absolute dream house.

6
00:00:27.839 --> 00:00:29.480
<v Speaker 3>Oh nice, Yeah, you.

7
00:00:29.359 --> 00:00:31.600
<v Speaker 2>Have the down payment saved up, your credit history is

8
00:00:31.960 --> 00:00:35.240
<v Speaker 2>completely solid, and you fill out the mortgage paperwork feeling

9
00:00:35.280 --> 00:00:39.920
<v Speaker 2>pretty confident, naturally, or picture applying for that perfect job

10
00:00:39.960 --> 00:00:44.200
<v Speaker 2>you're undeniably qualified for. You hits a bit, and almost instantly,

11
00:00:44.560 --> 00:00:48.119
<v Speaker 2>an automated email arrives, just saying application denied.

12
00:00:48.359 --> 00:00:50.479
<v Speaker 3>Ouch. Yeah, that's brutal right.

13
00:00:50.640 --> 00:00:53.520
<v Speaker 2>So you call the bank or the HR department and

14
00:00:53.560 --> 00:00:55.399
<v Speaker 2>you ask a very simple question, You just.

15
00:00:55.320 --> 00:00:58.520
<v Speaker 3>Ask why, good luck getting an answer to that exactly.

16
00:00:58.840 --> 00:01:01.119
<v Speaker 2>The person on the other end of the they hesitate.

17
00:01:01.159 --> 00:01:03.320
<v Speaker 2>They pull up your file and they say, we don't

18
00:01:03.399 --> 00:01:06.159
<v Speaker 2>actually know. The algorithm just flagged you as a high risk.

19
00:01:06.640 --> 00:01:09.799
<v Speaker 3>It is a uniquely modern kind of bureaucratic nightmare.

20
00:01:10.000 --> 00:01:11.159
<v Speaker 2>Really, it totally is.

21
00:01:11.319 --> 00:01:15.519
<v Speaker 3>Because there's no human reasoning to appeal to. There's no

22
00:01:15.519 --> 00:01:19.760
<v Speaker 3>no supervisor to escalate too. It's just this silent mathematical wall.

23
00:01:19.920 --> 00:01:22.239
<v Speaker 3>You are just trapped by the output of a system

24
00:01:22.280 --> 00:01:24.640
<v Speaker 3>that even its creators cannot fully explain.

25
00:01:24.959 --> 00:01:28.159
<v Speaker 2>Okay, let's unpack this because we are all basically living

26
00:01:28.159 --> 00:01:32.599
<v Speaker 2>with this black box problem in modern artificial intelligence. It's

27
00:01:32.640 --> 00:01:36.560
<v Speaker 2>a reality where these systems governing the trajectory of human

28
00:01:36.599 --> 00:01:38.840
<v Speaker 2>lives just operate in total.

29
00:01:38.560 --> 00:01:40.400
<v Speaker 3>Opacity, right completely in the dark.

30
00:01:40.640 --> 00:01:44.480
<v Speaker 2>But today we are going to explore this groundbreaking approach

31
00:01:44.640 --> 00:01:48.799
<v Speaker 2>engineered by researchers at Osaka Metropolitan University. It was led

32
00:01:48.799 --> 00:01:53.000
<v Speaker 2>by Professor yusikin Ojima, and they have essentially designed a

33
00:01:53.040 --> 00:01:55.680
<v Speaker 2>way out of the black box, which is huge. It's massive.

34
00:01:55.719 --> 00:01:58.159
<v Speaker 2>They aren't just trying to like slap a post trading

35
00:01:58.200 --> 00:02:01.719
<v Speaker 2>bandage on a biased neural network. They are building AI

36
00:02:01.959 --> 00:02:06.560
<v Speaker 2>that is inherently, fair fully auditible, and highly accurate, baked

37
00:02:06.680 --> 00:02:08.680
<v Speaker 2>right into the foundational architecture from day one.

38
00:02:08.879 --> 00:02:12.039
<v Speaker 3>Yeah, and to understand why this foundational shift is so necessary,

39
00:02:12.280 --> 00:02:14.560
<v Speaker 3>we really have to look at the architectural flaw in

40
00:02:14.639 --> 00:02:18.280
<v Speaker 3>how traditional deep learning models process the data we feed them.

41
00:02:18.360 --> 00:02:20.599
<v Speaker 2>Right, the data is the core issue exactly.

42
00:02:20.919 --> 00:02:26.599
<v Speaker 3>These algorithms they learn by consuming massive historical data sets,

43
00:02:27.080 --> 00:02:30.599
<v Speaker 3>so when a conventional AI analyzes, say, decades of past

44
00:02:30.680 --> 00:02:33.960
<v Speaker 3>loan approvals or hiring decisions, it doesn't just learn the

45
00:02:34.000 --> 00:02:35.520
<v Speaker 3>isolated markers of a good.

46
00:02:35.319 --> 00:02:37.360
<v Speaker 2>Candidate, because it's reading everything right.

47
00:02:37.599 --> 00:02:41.280
<v Speaker 3>It absorbs all the historical systemic biases related to race

48
00:02:41.439 --> 00:02:44.800
<v Speaker 3>or gender, age, income, all that stuff embedded in the data,

49
00:02:45.319 --> 00:02:47.840
<v Speaker 3>and it codifies them into its predictive models.

50
00:02:47.960 --> 00:02:50.120
<v Speaker 2>So the system essentially just acts as a mirror, right,

51
00:02:50.159 --> 00:02:52.960
<v Speaker 2>reflecting our historical flaws right back at us, but under

52
00:02:53.000 --> 00:02:56.120
<v Speaker 2>this guise of like objective mathematical certainty.

53
00:02:56.280 --> 00:02:58.599
<v Speaker 3>Yeah, the math makes it feel neutral when it really isn't.

54
00:02:58.719 --> 00:03:02.479
<v Speaker 2>And the real crisis, though, is the unexplainability factor. Because

55
00:03:02.479 --> 00:03:05.680
<v Speaker 2>once a deep learning model is fully trained, the logic

56
00:03:05.719 --> 00:03:09.319
<v Speaker 2>it uses to make a decision is distributed across billions,

57
00:03:09.439 --> 00:03:13.000
<v Speaker 2>sometimes literally trillions, of microscopic weights and parameters.

58
00:03:13.240 --> 00:03:16.479
<v Speaker 3>You're basically looking at a hyperdimensional matrix at that point.

59
00:03:16.319 --> 00:03:18.639
<v Speaker 2>Right, Because the reasoning is locked away in that dense

60
00:03:18.680 --> 00:03:21.520
<v Speaker 2>web of math. You can't just query the network and

61
00:03:21.560 --> 00:03:24.360
<v Speaker 2>to ask if it discriminated against a specific demographic it

62
00:03:24.439 --> 00:03:25.439
<v Speaker 2>just says no.

63
00:03:25.560 --> 00:03:28.960
<v Speaker 3>Which means proving systemic bias in a deep learning model

64
00:03:29.080 --> 00:03:32.199
<v Speaker 3>is often just an exercise and futility. I mean, you

65
00:03:32.199 --> 00:03:34.960
<v Speaker 3>can test the inputs and observe the outputs, but the

66
00:03:35.039 --> 00:03:37.400
<v Speaker 3>hidden layers remain completely opique.

67
00:03:37.599 --> 00:03:39.479
<v Speaker 2>So what does the industry usually do about this?

68
00:03:40.120 --> 00:03:43.240
<v Speaker 3>Well, the standard response has been to attempt these post

69
00:03:43.240 --> 00:03:47.719
<v Speaker 3>training fixes, so trying to mathematically adjust the final outputs

70
00:03:47.719 --> 00:03:51.199
<v Speaker 3>to enforce fairness metrics after the black box has already

71
00:03:51.199 --> 00:03:52.159
<v Speaker 3>done its processing.

72
00:03:52.599 --> 00:03:55.319
<v Speaker 2>I mean, it's like baking a cake with spoiled milk

73
00:03:55.319 --> 00:03:57.800
<v Speaker 2>and then trying to scrape off the bad taste with frosting.

74
00:03:58.120 --> 00:04:01.039
<v Speaker 3>That is a very accurate, if slightly gross, way to

75
00:04:01.039 --> 00:04:01.319
<v Speaker 3>put it.

76
00:04:01.439 --> 00:04:03.479
<v Speaker 2>Yes, you might make the outside look a little better,

77
00:04:03.639 --> 00:04:06.719
<v Speaker 2>but the underlying structural failure is still right there in

78
00:04:06.759 --> 00:04:10.439
<v Speaker 2>the cake. These post hawk adjustments almost always degrade the

79
00:04:10.479 --> 00:04:14.680
<v Speaker 2>system's overall predictive power, and they just fail entirely to

80
00:04:14.719 --> 00:04:17.839
<v Speaker 2>fix the root cause of the bias locked inside the model.

81
00:04:18.079 --> 00:04:20.800
<v Speaker 3>This race is an important question though. If patching the

82
00:04:20.839 --> 00:04:23.920
<v Speaker 3>AI doesn't work, how do we build fairness into the

83
00:04:23.920 --> 00:04:25.360
<v Speaker 3>foundation exactly?

84
00:04:25.800 --> 00:04:28.680
<v Speaker 2>We have to rethink the architecture itself. We have to

85
00:04:28.720 --> 00:04:32.399
<v Speaker 2>move away from rigid, opaque structures, which brings us to

86
00:04:32.480 --> 00:04:36.639
<v Speaker 2>the core of the Osaka team's innovation. They are shifting

87
00:04:36.639 --> 00:04:39.160
<v Speaker 2>the paradigm toward fuzzy logic.

88
00:04:39.319 --> 00:04:43.800
<v Speaker 3>Fuzzy logic, yeah, because traditional computing, and by extension, traditional

89
00:04:43.839 --> 00:04:48.720
<v Speaker 3>algorithmic decision making, it's built on strict Boolem.

90
00:04:48.160 --> 00:04:50.399
<v Speaker 2>Presholds, just ones and zeros.

91
00:04:50.160 --> 00:04:52.519
<v Speaker 3>Exactly yes or no. If a bank sets a hard

92
00:04:52.519 --> 00:04:54.319
<v Speaker 3>cut off for a prime interest rate at a credit

93
00:04:54.360 --> 00:04:57.120
<v Speaker 3>score of seven hundred, that is an absolute boundary. An

94
00:04:57.160 --> 00:04:58.800
<v Speaker 3>applicant with a seven hundred gets.

95
00:04:58.720 --> 00:05:00.639
<v Speaker 2>The primary and someone with a six ninety nine.

96
00:05:00.680 --> 00:05:03.480
<v Speaker 3>An applicant with a six ninety nine is automatically categorized

97
00:05:03.600 --> 00:05:06.439
<v Speaker 3>entirely differently, they might be denied or given a subprime rate.

98
00:05:06.560 --> 00:05:10.000
<v Speaker 2>So you cross this invisible, totally arbitrary line and your

99
00:05:10.040 --> 00:05:12.800
<v Speaker 2>outcome flips completely. It's like a simple on off light switch.

100
00:05:13.279 --> 00:05:16.680
<v Speaker 2>That kind of rigid categorization is practically the definition of

101
00:05:16.680 --> 00:05:20.079
<v Speaker 2>systemic unfairness because it ignores the reality that a six'

102
00:05:20.240 --> 00:05:23.839
<v Speaker 2>ninety nine and a seven hundred represent virtually identical financial

103
00:05:23.920 --> 00:05:24.839
<v Speaker 2>behaviors in the real.

104
00:05:24.879 --> 00:05:28.800
<v Speaker 3>World, right human beings rarely fit into perfect binary, Boxes.

105
00:05:28.519 --> 00:05:30.639
<v Speaker 2>So fuzzy logic is more like a dimmer. Switch, right

106
00:05:30.680 --> 00:05:32.360
<v Speaker 2>he steps in to smooth out that drop.

107
00:05:32.399 --> 00:05:34.759
<v Speaker 3>Off, yes a dimmer switch is the perfect way to

108
00:05:34.759 --> 00:05:38.040
<v Speaker 3>think about. It instead of a boolean, threshold fuzzy logic

109
00:05:38.079 --> 00:05:41.839
<v Speaker 3>maps these variables onto a continuous mathematical. Curve it uses

110
00:05:41.879 --> 00:05:43.439
<v Speaker 3>what are called degrees of.

111
00:05:43.560 --> 00:05:45.600
<v Speaker 2>Membership degrees of, membership, okay how does that?

112
00:05:45.639 --> 00:05:49.000
<v Speaker 3>Work so it assigns a degree of truth between zero and.

113
00:05:49.040 --> 00:05:53.839
<v Speaker 3>One rather than rigidly categorizing our applicant as just, unqualified

114
00:05:54.199 --> 00:05:57.759
<v Speaker 3>a fuzzy system might evaluate their score as having say

115
00:05:58.040 --> 00:06:00.839
<v Speaker 3>a point eight degree of membership in the good credit

116
00:06:00.920 --> 00:06:03.959
<v Speaker 3>category and a two degree in the average credit.

117
00:06:04.000 --> 00:06:05.519
<v Speaker 2>Category OH i, See.

118
00:06:05.319 --> 00:06:08.519
<v Speaker 3>And the rules use plain linguistic. Terms they take the

119
00:06:08.519 --> 00:06:11.199
<v Speaker 3>form of things like if income is mostly high and

120
00:06:11.319 --> 00:06:14.240
<v Speaker 3>debt is somewhat, low then the approval chance is very.

121
00:06:14.360 --> 00:06:17.800
<v Speaker 2>High that makes so much more. Sense by processing data

122
00:06:17.839 --> 00:06:21.120
<v Speaker 2>through degrees of, truth the system becomes incredibly robust against

123
00:06:21.160 --> 00:06:24.839
<v Speaker 2>those arbitrary boundary. Issues so a minor perturbation in the,

124
00:06:24.879 --> 00:06:28.199
<v Speaker 2>input like making one dollar less than a, threshold or

125
00:06:28.240 --> 00:06:30.560
<v Speaker 2>scoring a fraction of a point lower on some, assessment

126
00:06:30.920 --> 00:06:32.920
<v Speaker 2>it doesn't trigger a massive disproportionate.

127
00:06:32.920 --> 00:06:36.759
<v Speaker 3>Penalty, exactly the transition is. Smooth small changes in input

128
00:06:36.879 --> 00:06:39.199
<v Speaker 3>don't make decisions flip, abruptly.

129
00:06:39.079 --> 00:06:45.079
<v Speaker 2>Which means similar individuals receive mathematically similar. Evaluations that establishes

130
00:06:45.079 --> 00:06:48.079
<v Speaker 2>a baseline of structural equity before we even get to

131
00:06:48.079 --> 00:06:49.199
<v Speaker 2>the complex modeling.

132
00:06:49.240 --> 00:06:53.959
<v Speaker 3>Part it, does but that smooth transition introduces a massive engineering.

133
00:06:54.040 --> 00:06:56.800
<v Speaker 3>CHALLENGE i, mean it is one thing to manually write

134
00:06:56.839 --> 00:06:59.319
<v Speaker 3>a dozen fuzzy rules to control the temperature of an air,

135
00:06:59.319 --> 00:07:03.560
<v Speaker 3>Conditioner but it is an entirely different universe to write

136
00:07:03.600 --> 00:07:07.800
<v Speaker 3>the millions of, overlapping nuanced fuzzy rules required to evaluate

137
00:07:07.839 --> 00:07:10.800
<v Speaker 3>the global credit market or a modern hiring.

138
00:07:10.879 --> 00:07:13.959
<v Speaker 2>Pipeline, yeah a human programmer just can't sit down and

139
00:07:14.040 --> 00:07:17.040
<v Speaker 2>manually code all the linguistic rules for a data set

140
00:07:17.199 --> 00:07:18.560
<v Speaker 2>with hundreds of intersecting.

141
00:07:18.639 --> 00:07:20.160
<v Speaker 3>Variables no, Way it's too.

142
00:07:20.240 --> 00:07:23.240
<v Speaker 2>Complex so if human engineering can't scale to meet, it

143
00:07:23.560 --> 00:07:26.759
<v Speaker 2>and deep neural networks are too opaque to, trust we

144
00:07:26.800 --> 00:07:29.920
<v Speaker 2>need an algorithm that can autonomously write and optimize its

145
00:07:29.920 --> 00:07:30.639
<v Speaker 2>own transparent.

146
00:07:30.720 --> 00:07:33.439
<v Speaker 3>Logic we need the system to build itself.

147
00:07:33.319 --> 00:07:36.399
<v Speaker 2>Right and that leads us directly to the fascinating. Mechanism

148
00:07:36.480 --> 00:07:39.279
<v Speaker 2>The osaka team utilized genetics based machine.

149
00:07:39.319 --> 00:07:43.160
<v Speaker 3>Learning, yeah applying the principles Of darwinian evolution to software.

150
00:07:43.240 --> 00:07:48.040
<v Speaker 2>Architecture it is so. Cool the algorithm generates these human

151
00:07:48.079 --> 00:07:50.480
<v Speaker 2>readable fuzzy rules through a process of.

152
00:07:50.480 --> 00:07:52.800
<v Speaker 3>Natural selection survival of the, fittest.

153
00:07:52.519 --> 00:07:56.040
<v Speaker 2>Exactly survival of the. Fittest it creates populations of rule,

154
00:07:56.120 --> 00:08:00.360
<v Speaker 2>sets and the better performing combinations survive and, reproduce while

155
00:08:00.360 --> 00:08:02.560
<v Speaker 2>the weaker ones are just, discarded.

156
00:08:02.040 --> 00:08:05.000
<v Speaker 3>And it literally splices the code. Together it might take

157
00:08:05.000 --> 00:08:08.920
<v Speaker 3>the if condition from one highly successful rule and combine

158
00:08:08.959 --> 00:08:11.759
<v Speaker 3>it with the then outcome of. Another it performs this

159
00:08:11.879 --> 00:08:15.519
<v Speaker 3>mathematical crossover to create a new generation of hybrid.

160
00:08:15.600 --> 00:08:15.920
<v Speaker 2>Rules.

161
00:08:16.040 --> 00:08:19.680
<v Speaker 3>Wow and it also introduces random, mutations you, know slightly

162
00:08:19.720 --> 00:08:22.399
<v Speaker 3>altering a threshold or a linguistic, variable just to make

163
00:08:22.399 --> 00:08:24.959
<v Speaker 3>sure the system continually explores new logical.

164
00:08:25.000 --> 00:08:28.360
<v Speaker 2>Pathways but, wait if we force THE ai to care

165
00:08:28.439 --> 00:08:33.000
<v Speaker 2>about fairness and, simplicity doesn't the accuracy completely? Tank you

166
00:08:33.039 --> 00:08:34.320
<v Speaker 2>can't have your cake and eat it, too.

167
00:08:34.399 --> 00:08:37.480
<v Speaker 3>Right that is the exact technical hurdle that has stalled

168
00:08:37.519 --> 00:08:42.039
<v Speaker 3>EQUITABLE ai for, years the assumption that fairness inherently destroys.

169
00:08:42.080 --> 00:08:45.799
<v Speaker 3>Performance but the evolutionary process fundamentally challenges.

170
00:08:45.799 --> 00:08:49.360
<v Speaker 2>That how so because usually if you optimize for multiple competing,

171
00:08:49.440 --> 00:08:51.519
<v Speaker 2>things you just get, mediocrity.

172
00:08:51.120 --> 00:08:55.159
<v Speaker 3>Right but they use multi objective optimization THE ai isn't

173
00:08:55.200 --> 00:08:57.960
<v Speaker 3>just trying to evolve to be the most. Accurate it

174
00:08:58.000 --> 00:09:03.279
<v Speaker 3>is simultaneously optimizing for three things overall, accuracy, fairness metrics

175
00:09:03.279 --> 00:09:07.360
<v Speaker 3>across demographic, groups and. Interpretability so keeping the rules simple

176
00:09:07.639 --> 00:09:09.360
<v Speaker 3>and few in, number, Okay.

177
00:09:09.120 --> 00:09:11.879
<v Speaker 2>But how does it balance all three without just failing

178
00:09:11.879 --> 00:09:12.279
<v Speaker 2>at all of?

179
00:09:12.320 --> 00:09:15.279
<v Speaker 3>Them by exploring what is known as The peretto. Front

180
00:09:15.639 --> 00:09:18.879
<v Speaker 3>because the genetic algorithm explores a vast search space over

181
00:09:18.960 --> 00:09:23.080
<v Speaker 3>thousands of, generations it doesn't just spit out one compromise rule.

182
00:09:23.120 --> 00:09:26.200
<v Speaker 3>Set it maps an entire curve of optimal trade off.

183
00:09:26.360 --> 00:09:28.440
<v Speaker 2>Solutions, ah so it gives you options.

184
00:09:28.639 --> 00:09:32.600
<v Speaker 3>Exactly in their experiments with real world benchmark data, sets

185
00:09:32.720 --> 00:09:35.600
<v Speaker 3>they found that a, tiny highly controlled reduction in overall

186
00:09:35.639 --> 00:09:38.519
<v Speaker 3>accuracy often led to massive sweeping gains in.

187
00:09:38.559 --> 00:09:40.480
<v Speaker 2>Fairness, really so you just give up a little bit

188
00:09:40.519 --> 00:09:42.399
<v Speaker 2>of perfection for a lot of equity.

189
00:09:42.200 --> 00:09:45.960
<v Speaker 3>Exactly and sometimes the fuzzy rules even improved both metrics.

190
00:09:46.000 --> 00:09:48.960
<v Speaker 2>Simultaneously, wait it got more accurate and more fair at

191
00:09:49.000 --> 00:09:49.279
<v Speaker 2>the same.

192
00:09:49.360 --> 00:09:54.240
<v Speaker 3>Time, Yeah by finding nuanced nonlinear relationships in the data

193
00:09:54.279 --> 00:09:58.240
<v Speaker 3>that the rigid binary deep learning models completely, missed it

194
00:09:58.320 --> 00:09:59.679
<v Speaker 3>stops penalizing the edge.

195
00:09:59.679 --> 00:10:02.440
<v Speaker 2>Cases, yes that is, incredible but how does it actually

196
00:10:02.519 --> 00:10:04.440
<v Speaker 2>know it's being? Fair, like how does it measure that

197
00:10:04.519 --> 00:10:04.799
<v Speaker 2>during the?

198
00:10:04.840 --> 00:10:07.399
<v Speaker 3>Evolution it checks its own fairness from the very first

199
00:10:07.399 --> 00:10:14.360
<v Speaker 3>generation using standard verifiable, indicators things like demographic parity and equal, opportunity.

200
00:10:13.840 --> 00:10:18.120
<v Speaker 2>Meaning it ensures positive outcomes are distributed evenly regardless of

201
00:10:18.399 --> 00:10:20.039
<v Speaker 2>sensitive attributes like race or.

202
00:10:20.080 --> 00:10:24.480
<v Speaker 3>Gender, exactly it mathematically guarantees that positive outcomes aren't. Skewed

203
00:10:25.080 --> 00:10:28.600
<v Speaker 3>the code literally cannot survive the evolutionary process if it

204
00:10:28.639 --> 00:10:29.559
<v Speaker 3>relies on biased.

205
00:10:29.639 --> 00:10:33.519
<v Speaker 2>Logic this changes everything for high stakes. Applications just think about,

206
00:10:33.600 --> 00:10:37.480
<v Speaker 2>hiring but evaluating candidates and ensuring similar qualifications means similar

207
00:10:37.519 --> 00:10:40.519
<v Speaker 2>interview chances regardless of demographics or.

208
00:10:40.519 --> 00:10:44.759
<v Speaker 3>Lending preventing disproportionate loan denials for specific, neighborhoods you, know

209
00:10:44.799 --> 00:10:46.759
<v Speaker 3>avoiding digital, redlining.

210
00:10:46.440 --> 00:10:50.799
<v Speaker 2>Or even healthcare recommending treatments without biases and, diagnostics and

211
00:10:50.919 --> 00:10:54.720
<v Speaker 2>criminal justice balancing public safety with equitable risk assessments for

212
00:10:54.759 --> 00:10:55.200
<v Speaker 2>bail or.

213
00:10:55.240 --> 00:10:58.159
<v Speaker 3>Parole the stakes are incredibly high in all those.

214
00:10:58.200 --> 00:11:01.360
<v Speaker 2>Fields here's where it gets really, interesting, though the audit.

215
00:11:01.440 --> 00:11:05.960
<v Speaker 2>Power because these fuzzy models use dozens of rules instead

216
00:11:05.960 --> 00:11:08.360
<v Speaker 2>of thousands or, millions humans can actually read.

217
00:11:08.399 --> 00:11:11.799
<v Speaker 3>Them that's the interpretability constraint kicking. In it's a massive

218
00:11:11.840 --> 00:11:12.840
<v Speaker 3>advantage over black.

219
00:11:12.879 --> 00:11:17.440
<v Speaker 2>Boxes it's. Huge this means, regulators, ethics, boards and even

220
00:11:17.480 --> 00:11:21.200
<v Speaker 2>everyday people can trace a decision back to the exact

221
00:11:21.759 --> 00:11:23.279
<v Speaker 2>human readable rule that caused.

222
00:11:23.320 --> 00:11:25.360
<v Speaker 3>It, right you can point to a rule and say

223
00:11:25.480 --> 00:11:26.240
<v Speaker 3>this is why you were.

224
00:11:26.279 --> 00:11:30.120
<v Speaker 2>Denied and that built in auditability builds crucial public. Trust

225
00:11:30.799 --> 00:11:34.120
<v Speaker 2>when the logic is laid, bare society can actually debate

226
00:11:34.159 --> 00:11:37.840
<v Speaker 2>the validity of the rules, themselves rather than just fearing

227
00:11:37.879 --> 00:11:41.240
<v Speaker 2>the hidden machinations of a. Machine so, True so summarizing

228
00:11:41.279 --> 00:11:43.799
<v Speaker 2>the big picture, here fairness does not have to come

229
00:11:43.879 --> 00:11:48.000
<v Speaker 2>at the expensive. USEFULNESS ai can deliver accurate predictions while

230
00:11:48.039 --> 00:11:50.919
<v Speaker 2>treating people equitably and giving clear reasons for its.

231
00:11:51.039 --> 00:11:53.440
<v Speaker 3>Choices and if we connect this to the bigger, picture

232
00:11:53.879 --> 00:11:57.320
<v Speaker 3>the future applications are just. Vast future work could extend

233
00:11:57.360 --> 00:12:00.879
<v Speaker 3>this to dynamic environments where fairness needs change time.

234
00:12:00.879 --> 00:12:03.720
<v Speaker 2>Oh like as the economy shifts the rules.

235
00:12:03.399 --> 00:12:07.240
<v Speaker 3>Adapt, exactly or integrating it with federated learning where the

236
00:12:07.360 --> 00:12:11.039
<v Speaker 3>data stays distributed across different organizations like multiple hospitals for,

237
00:12:11.159 --> 00:12:14.879
<v Speaker 3>privacy but the algorithm still learns from all of.

238
00:12:14.919 --> 00:12:16.639
<v Speaker 2>It that would be amazing for medical.

239
00:12:16.679 --> 00:12:20.840
<v Speaker 3>Research, right there's even potential to hybridize these fuzzy systems

240
00:12:21.120 --> 00:12:25.120
<v Speaker 3>with larger neural, networks creating incredibly powerful but transparent.

241
00:12:25.200 --> 00:12:29.840
<v Speaker 2>Solutions it really reframes our entire relationship with. Algorithms we

242
00:12:29.919 --> 00:12:33.759
<v Speaker 2>actually have the technology now to MAKE ai perfectly transparent

243
00:12:33.799 --> 00:12:35.840
<v Speaker 2>and fair by human, standards.

244
00:12:35.399 --> 00:12:37.960
<v Speaker 3>Which is an incredible milestone.

245
00:12:37.279 --> 00:12:39.759
<v Speaker 2>It, is but it leads us with a lingering question to.

246
00:12:39.799 --> 00:12:43.679
<v Speaker 2>Ponder since we can now perfectly program THE ai to

247
00:12:43.799 --> 00:12:47.320
<v Speaker 2>follow our definitions of, equity who gets to decide which

248
00:12:47.360 --> 00:12:50.559
<v Speaker 2>definition of fairness THE ai should evolve? Toward that is

249
00:12:50.600 --> 00:12:53.320
<v Speaker 2>the real, Challenge, Right if THE a could balance the scales,

250
00:12:53.360 --> 00:12:56.519
<v Speaker 2>perfectly the hardest part might just be humans agreeing on

251
00:12:56.559 --> 00:12:58.200
<v Speaker 2>what a balanced scale actually looks. Like