How I'm fighting bias in algorithms | Joy Buolamwini

308,224 views ・ 2017-03-29

TED


Please double-click on the English subtitles below to play the video.

00:12
Hello, I'm Joy, a poet of code,
0
12861
3134
00:16
on a mission to stop an unseen force that's rising,
1
16019
4993
00:21
a force that I called "the coded gaze,"
2
21036
2856
00:23
my term for algorithmic bias.
3
23916
3309
00:27
Algorithmic bias, like human bias, results in unfairness.
4
27249
4300
00:31
However, algorithms, like viruses, can spread bias on a massive scale
5
31573
6022
00:37
at a rapid pace.
6
37619
1582
00:39
Algorithmic bias can also lead to exclusionary experiences
7
39763
4387
00:44
and discriminatory practices.
8
44174
2128
00:46
Let me show you what I mean.
9
46326
2061
00:48
(Video) Joy Buolamwini: Hi, camera. I've got a face.
10
48800
2436
00:51
Can you see my face?
11
51982
1864
00:53
No-glasses face?
12
53871
1625
00:55
You can see her face.
13
55521
2214
00:58
What about my face?
14
58057
2245
01:03
I've got a mask. Can you see my mask?
15
63710
3750
01:08
Joy Buolamwini: So how did this happen?
16
68294
2365
01:10
Why am I sitting in front of a computer
17
70683
3141
01:13
in a white mask,
18
73848
1424
01:15
trying to be detected by a cheap webcam?
19
75296
3650
01:18
Well, when I'm not fighting the coded gaze
20
78970
2291
01:21
as a poet of code,
21
81285
1520
01:22
I'm a graduate student at the MIT Media Lab,
22
82829
3272
01:26
and there I have the opportunity to work on all sorts of whimsical projects,
23
86125
4917
01:31
including the Aspire Mirror,
24
91066
2027
01:33
a project I did so I could project digital masks onto my reflection.
25
93117
5134
01:38
So in the morning, if I wanted to feel powerful,
26
98275
2350
01:40
I could put on a lion.
27
100649
1434
01:42
If I wanted to be uplifted, I might have a quote.
28
102107
3496
01:45
So I used generic facial recognition software
29
105627
2989
01:48
to build the system,
30
108640
1351
01:50
but found it was really hard to test it unless I wore a white mask.
31
110015
5103
01:56
Unfortunately, I've run into this issue before.
32
116102
4346
02:00
When I was an undergraduate at Georgia Tech studying computer science,
33
120472
4303
02:04
I used to work on social robots,
34
124799
2055
02:06
and one of my tasks was to get a robot to play peek-a-boo,
35
126878
3777
02:10
a simple turn-taking game
36
130679
1683
02:12
where partners cover their face and then uncover it saying, "Peek-a-boo!"
37
132386
4321
02:16
The problem is, peek-a-boo doesn't really work if I can't see you,
38
136731
4429
02:21
and my robot couldn't see me.
39
141184
2499
02:23
But I borrowed my roommate's face to get the project done,
40
143707
3950
02:27
submitted the assignment,
41
147681
1380
02:29
and figured, you know what, somebody else will solve this problem.
42
149085
3753
02:33
Not too long after,
43
153489
2003
02:35
I was in Hong Kong for an entrepreneurship competition.
44
155516
4159
02:40
The organizers decided to take participants
45
160159
2694
02:42
on a tour of local start-ups.
46
162877
2372
02:45
One of the start-ups had a social robot,
47
165273
2715
02:48
and they decided to do a demo.
48
168012
1912
02:49
The demo worked on everybody until it got to me,
49
169948
2980
02:52
and you can probably guess it.
50
172952
1923
02:54
It couldn't detect my face.
51
174899
2965
02:57
I asked the developers what was going on,
52
177888
2511
03:00
and it turned out we had used the same generic facial recognition software.
53
180423
5533
03:05
Halfway around the world,
54
185980
1650
03:07
I learned that algorithmic bias can travel as quickly
55
187654
3852
03:11
as it takes to download some files off of the internet.
56
191530
3170
03:15
So what's going on? Why isn't my face being detected?
57
195565
3076
03:18
Well, we have to look at how we give machines sight.
58
198665
3356
03:22
Computer vision uses machine learning techniques
59
202045
3409
03:25
to do facial recognition.
60
205478
1880
03:27
So how this works is, you create a training set with examples of faces.
61
207382
3897
03:31
This is a face. This is a face. This is not a face.
62
211303
2818
03:34
And over time, you can teach a computer how to recognize other faces.
63
214145
4519
03:38
However, if the training sets aren't really that diverse,
64
218688
3989
03:42
any face that deviates too much from the established norm
65
222701
3349
03:46
will be harder to detect,
66
226074
1649
03:47
which is what was happening to me.
67
227747
1963
03:49
But don't worry -- there's some good news.
68
229734
2382
03:52
Training sets don't just materialize out of nowhere.
69
232140
2771
03:54
We actually can create them.
70
234935
1788
03:56
So there's an opportunity to create full-spectrum training sets
71
236747
4176
04:00
that reflect a richer portrait of humanity.
72
240947
3824
04:04
Now you've seen in my examples
73
244795
2221
04:07
how social robots
74
247040
1768
04:08
was how I found out about exclusion with algorithmic bias.
75
248832
4611
04:13
But algorithmic bias can also lead to discriminatory practices.
76
253467
4815
04:19
Across the US,
77
259257
1453
04:20
police departments are starting to use facial recognition software
78
260734
4198
04:24
in their crime-fighting arsenal.
79
264956
2459
04:27
Georgetown Law published a report
80
267439
2013
04:29
showing that one in two adults in the US -- that's 117 million people --
81
269476
6763
04:36
have their faces in facial recognition networks.
82
276263
3534
04:39
Police departments can currently look at these networks unregulated,
83
279821
4552
04:44
using algorithms that have not been audited for accuracy.
84
284397
4286
04:48
Yet we know facial recognition is not fail proof,
85
288707
3864
04:52
and labeling faces consistently remains a challenge.
86
292595
4179
04:56
You might have seen this on Facebook.
87
296798
1762
04:58
My friends and I laugh all the time when we see other people
88
298584
2988
05:01
mislabeled in our photos.
89
301596
2458
05:04
But misidentifying a suspected criminal is no laughing matter,
90
304078
5591
05:09
nor is breaching civil liberties.
91
309693
2827
05:12
Machine learning is being used for facial recognition,
92
312544
3205
05:15
but it's also extending beyond the realm of computer vision.
93
315773
4505
05:21
In her book, "Weapons of Math Destruction,"
94
321086
4016
05:25
data scientist Cathy O'Neil talks about the rising new WMDs --
95
325126
6681
05:31
widespread, mysterious and destructive algorithms
96
331831
4353
05:36
that are increasingly being used to make decisions
97
336208
2964
05:39
that impact more aspects of our lives.
98
339196
3177
05:42
So who gets hired or fired?
99
342397
1870
05:44
Do you get that loan? Do you get insurance?
100
344291
2112
05:46
Are you admitted into the college you wanted to get into?
101
346427
3503
05:49
Do you and I pay the same price for the same product
102
349954
3509
05:53
purchased on the same platform?
103
353487
2442
05:55
Law enforcement is also starting to use machine learning
104
355953
3759
05:59
for predictive policing.
105
359736
2289
06:02
Some judges use machine-generated risk scores to determine
106
362049
3494
06:05
how long an individual is going to spend in prison.
107
365567
4402
06:09
So we really have to think about these decisions.
108
369993
2454
06:12
Are they fair?
109
372471
1182
06:13
And we've seen that algorithmic bias
110
373677
2890
06:16
doesn't necessarily always lead to fair outcomes.
111
376591
3374
06:19
So what can we do about it?
112
379989
1964
06:21
Well, we can start thinking about how we create more inclusive code
113
381977
3680
06:25
and employ inclusive coding practices.
114
385681
2990
06:28
It really starts with people.
115
388695
2309
06:31
So who codes matters.
116
391528
1961
06:33
Are we creating full-spectrum teams with diverse individuals
117
393513
4119
06:37
who can check each other's blind spots?
118
397656
2411
06:40
On the technical side, how we code matters.
119
400091
3545
06:43
Are we factoring in fairness as we're developing systems?
120
403660
3651
06:47
And finally, why we code matters.
121
407335
2913
06:50
We've used tools of computational creation to unlock immense wealth.
122
410605
5083
06:55
We now have the opportunity to unlock even greater equality
123
415712
4447
07:00
if we make social change a priority
124
420183
2930
07:03
and not an afterthought.
125
423137
2170
07:05
And so these are the three tenets that will make up the "incoding" movement.
126
425828
4522
07:10
Who codes matters,
127
430374
1652
07:12
how we code matters
128
432050
1543
07:13
and why we code matters.
129
433617
2023
07:15
So to go towards incoding, we can start thinking about
130
435664
3099
07:18
building platforms that can identify bias
131
438787
3164
07:21
by collecting people's experiences like the ones I shared,
132
441975
3078
07:25
but also auditing existing software.
133
445077
3070
07:28
We can also start to create more inclusive training sets.
134
448171
3765
07:31
Imagine a "Selfies for Inclusion" campaign
135
451960
2803
07:34
where you and I can help developers test and create
136
454787
3655
07:38
more inclusive training sets.
137
458466
2093
07:41
And we can also start thinking more conscientiously
138
461122
2828
07:43
about the social impact of the technology that we're developing.
139
463974
5391
07:49
To get the incoding movement started,
140
469389
2393
07:51
I've launched the Algorithmic Justice League,
141
471806
2847
07:54
where anyone who cares about fairness can help fight the coded gaze.
142
474677
5872
08:00
On codedgaze.com, you can report bias,
143
480573
3296
08:03
request audits, become a tester
144
483893
2445
08:06
and join the ongoing conversation,
145
486362
2771
08:09
#codedgaze.
146
489157
2287
08:12
So I invite you to join me
147
492562
2487
08:15
in creating a world where technology works for all of us,
148
495073
3719
08:18
not just some of us,
149
498816
1897
08:20
a world where we value inclusion and center social change.
150
500737
4588
08:25
Thank you.
151
505349
1175
08:26
(Applause)
152
506548
4271
08:32
But I have one question:
153
512693
2854
08:35
Will you join me in the fight?
154
515571
2059
08:37
(Laughter)
155
517654
1285
08:38
(Applause)
156
518963
3687
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7