AI is Biased, But Do We Know Why? | Sadhana Lolla | TEDxBoston — Transcript

Full transcript

1. 0:08As a computer scientist and an
2. 0:10artificial intelligence researcher,
3. 0:12there's no doubt in my mind that AI is
4. 0:14going to change the fabric of our lives.
5. 0:17But, before AI can truly be deployed
6. 0:20into the areas where it is the most
7. 0:21necessary, like safety-critical domains,
8. 0:24there are still a number of key issues
9. 0:26that we, as engineers and researchers,
10. 0:29must address.
11. 0:31Let me give you an example using Chat
12. 0:33GPT, a really popular language model. We
13. 0:36can ask Chat GPT questions and it will
14. 0:38respond intelligently. And in this
15. 0:40example, we ask Chat GPT to list 10
16. 0:43philosophers, and it gives us a pretty
17. 0:45good list.
18. 0:47But, if you'll notice, all of the
19. 0:49individuals listed on the screen are
20. 0:51men.
21. 0:53So, when we ask Chat GPT why it didn't
22. 0:56list any women, it apologizes profusely,
23. 0:59which is very kind, and it lists 10
24. 1:02women philosophers.
25. 1:04And we can continue to ask Chat GPT
26. 1:06questions like this, asking it for
27. 1:09non-Western philosophers, non-Western
28. 1:11women philosophers, and once we have
29. 1:13this diverse list of philosophers, we
30. 1:16can go back and ask, "Okay, let's try
31. 1:18again.
32. 1:19Give me 10 philosophers."
33. 1:22And instead,
34. 1:24it gives us the same exact list of 10
35. 1:27white, Western, male philosophers.
36. 1:31So, on the surface, this seems like a
37. 1:33pretty trivial problem. Nobody was hurt
38. 1:36in the creation of this video. Um there
39. 1:38are no real consequences to Chat GPT
40. 1:40answering questions like this.
41. 1:42But in reality, this is a classic
42. 1:44example of algorithmic bias, which is
43. 1:47present in almost every single model
44. 1:50that has been deployed today.
45. 1:53So, when you and I think of bias, we're
46. 1:55often thinking about human bias, which
47. 1:57is when we are prejudiced towards a
48. 1:59group of individuals or a system of
49. 2:02beliefs.
50. 2:03But bias in AI, even though it may
51. 2:05propagate human biases, is completely
52. 2:08different. Because it is quantifiable.
53. 2:10We can assign a number to this, and more
54. 2:13importantly, using this type of
55. 2:14analysis, we can mitigate algorithmic
56. 2:17bias.
57. 2:18And that's because algorithmic bias
58. 2:20fundamentally results because of
59. 2:22imbalances in data that are used to
60. 2:24train artificial intelligence models.
61. 2:27So, here's an example, an oversimplified
62. 2:29one, of a data distribution that we
63. 2:31might use to train an AI system.
64. 2:34There are peaks in this distribution, so
65. 2:36areas of high data, and there are areas
66. 2:39of lower representation, places where we
67. 2:41don't have a lot of data.
68. 2:44When we're training models that involve
69. 2:46human data, the data that is at the
70. 2:48peaks of this distribution tends to come
71. 2:51from, exactly as we just saw,
72. 2:53white, Western men.
73. 2:56And data that comes from the
74. 2:57underrepresented regions of this type of
75. 2:59data set comes from women and people of
76. 3:02color.
77. 3:04So, what would happen if we trained an
78. 3:06artificial intelligence system on a data
79. 3:08set that looked something like this?
80. 3:11Let's consider the example of facial
81. 3:12detection.
82. 3:14Artificial intelligence-powered facial
83. 3:15detection systems are everywhere. We use
84. 3:18them on our phones daily. They're also
85. 3:20used in surveillance systems, and the
86. 3:22reports of these surveillance systems
87. 3:24are used in the criminal justice system.
88. 3:27But these algorithms are fundamentally
89. 3:29biased towards lighter-skinned faces
90. 3:32because the overrepresented areas of
91. 3:33this data set of a lot of commercial
92. 3:35facial detection systems is a lot of
93. 3:38lighter-skinned faces.
94. 3:40So, these algorithms end up being
95. 3:42infamously bad at distinguishing
96. 3:45darker-skinned faces from backgrounds,
97. 3:47but they're very good at picking out
98. 3:49lighter-skinned faces. And this can have
99. 3:51pretty catastrophic effects.
100. 3:54But, although bias in artificial
101. 3:55intelligence may propagate and amplify
102. 3:58human social and racial biases,
103. 4:00the problem with bias in artificial
104. 4:02intelligence is a much more foundational
105. 4:04one. It is due only to data
106. 4:07distributions.
107. 4:09And so, this can present itself in a lot
108. 4:11of different ways.
109. 4:13Let's consider the example of an AI
110. 4:15system that is using
111. 4:17um that is trying to predict whether or
112. 4:18not a human has a very rare disease.
113. 4:22In this case, the AI system is mostly
114. 4:24trained on healthy patients because
115. 4:26they're so much more common than
116. 4:28individuals who are sick or who have
117. 4:29this super rare disease.
118. 4:32And that means that the model is
119. 4:33fundamentally biased towards healthy
120. 4:35patients, and it won't actually be able
121. 4:37to determine which patients have a
122. 4:39disease. And this can lead to a lot of
123. 4:41false negatives, and that can have
124. 4:43life-altering consequences. So, I truly
125. 4:47believe that mitigating algorithmic bias
126. 4:49is one of the largest challenges to
127. 4:51modern artificial intelligence
128. 4:53because without this mitigation, we
129. 4:55won't be able to deploy any of the
130. 4:57advances that we currently have into the
131. 4:59real world because they won't be able to
132. 5:02perform well on everyone who needs to
133. 5:04use them. So, at MIT and at Themis AI,
134. 5:08which is um a startup that has spun out
135. 5:10from MIT, this is exactly the type of
136. 5:12problem that we are trying to address.
137. 5:15We have developed systematic algorithmic
138. 5:18approaches to debiasing artificial
139. 5:20intelligence algorithms,
140. 5:21even those where they're trained on such
141. 5:24imbalanced data sets.
142. 5:26And the way we do this is we fully
143. 5:28unpack how an artificial intelligence
144. 5:30algorithm is able to extract data from
145. 5:33input, and we make sure that this
146. 5:35algorithm is able to perform well on the
147. 5:38most challenging scenarios that we can
148. 5:40give it.
149. 5:41And this results in a new class of
150. 5:43models, AI systems that not only work on
151. 5:46some of the data in their data sets, but
152. 5:48all of them.
153. 5:51So, today I hope that you've learned a
154. 5:53little bit more about how bias in
155. 5:56artificial intelligence really works,
156. 5:58how the inputs to a model can affect how
157. 6:00its performance works on all of us.
158. 6:03And I truly believe that by mitigating
159. 6:05algorithmic bias and building
160. 6:07trustworthy bias-free AI, we can get one
161. 6:10step closer to deploying artificial
162. 6:12intelligence in safety-critical domains.
163. 6:16By creating these trustworthy artificial
164. 6:19intelligence algorithms, we're one step
165. 6:21closer to robotic surgeons operating in
166. 6:23hospitals, autonomous vehicles driving
167. 6:26on all of our roads, healthcare AI
168. 6:28systems that are actually able to
169. 6:29diagnose previously hidden diseases, and
170. 6:32we can make sure that all of these
171. 6:33advances work not just for some of us,
172. 6:36but for all of us. Thank you.