WEBVTT
1
00:00:01.577 --> 00:00:03.550
(soft music)
2
00:00:03.550 --> 00:00:05.940
Are algorithms getting less important?
3
00:00:05.940 --> 00:00:08.090
As algorithms become commoditized,
4
00:00:08.090 --> 00:00:09.800
it may be less about the algorithm
5
00:00:09.800 --> 00:00:11.350
and more about the application.
6
00:00:12.490 --> 00:00:16.320
In our first episode of season two of "Me, Myself, and AI",
7
00:00:16.320 --> 00:00:17.990
we'll talk with Craig Martel,
8
00:00:17.990 --> 00:00:19.580
Head of Machine Learning at Lyft
9
00:00:19.580 --> 00:00:21.930
about how Lyft uses artificial intelligence
10
00:00:21.930 --> 00:00:23.980
to improve its business.
11
00:00:23.980 --> 00:00:25.950
Welcome to "Me, Myself, and AI",
12
00:00:25.950 --> 00:00:29.180
a podcast on artificial intelligence in business.
13
00:00:29.180 --> 00:00:30.690
Each episode, we introduce you
14
00:00:30.690 --> 00:00:33.020
to someone innovating with AI.
15
00:00:33.020 --> 00:00:35.930
I'm Sam Ramsbotham, Professor of Information Systems
16
00:00:35.930 --> 00:00:37.570
at Boston College.
17
00:00:37.570 --> 00:00:39.597
I'm also the guest editor for the AI
18
00:00:39.597 --> 00:00:41.870
and business strategy big idea program,
19
00:00:41.870 --> 00:00:44.660
at MIT Sloan Management Review.
20
00:00:44.660 --> 00:00:46.810
And I'm Shervin Khodabandeh,
21
00:00:46.810 --> 00:00:48.510
Senior Partner with BCG
22
00:00:48.510 --> 00:00:52.360
and I co-lead BCG's AI practice in North America.
23
00:00:52.360 --> 00:00:55.250
And together MIT SMR and BCG
24
00:00:55.250 --> 00:00:57.860
have been researching AI for five years,
25
00:00:57.860 --> 00:01:00.040
interviewing hundreds of practitioners
26
00:01:00.040 --> 00:01:02.060
and surveying thousands of companies
27
00:01:02.060 --> 00:01:03.990
on what it takes to build
28
00:01:03.990 --> 00:01:07.180
and to deploy and scale AI capabilities across
29
00:01:07.180 --> 00:01:08.960
the organization and really transform
30
00:01:08.960 --> 00:01:10.593
the way organizations operate.
31
00:01:13.410 --> 00:01:15.320
Today we're talking with Craig Martel.
32
00:01:15.320 --> 00:01:17.720
Craig is the Head of Machine Learning for Lyft.
33
00:01:17.720 --> 00:01:19.480
Thanks for joining us today, Craig.
34
00:01:19.480 --> 00:01:20.980
Thanks Sam, I'm really happy to be here,
35
00:01:20.980 --> 00:01:22.760
these are pretty exciting topics.
36
00:01:22.760 --> 00:01:24.960
So Craig, Head of Machine Learning at Lyft,
37
00:01:24.960 --> 00:01:28.400
what exactly does that mean and how did you get there?
38
00:01:28.400 --> 00:01:30.470
So let me start by saying
39
00:01:30.470 --> 00:01:32.320
I'm pretty sure I won the lottery in life
40
00:01:32.320 --> 00:01:33.740
and here's why.
41
00:01:33.740 --> 00:01:36.790
I started off doing political theory academically
42
00:01:36.790 --> 00:01:39.280
and I have this misspent youth
43
00:01:39.280 --> 00:01:40.370
where I gathered a collection
44
00:01:40.370 --> 00:01:42.240
of master's degrees along the way
45
00:01:42.240 --> 00:01:43.780
to figuring out what I want to do.
46
00:01:43.780 --> 00:01:45.910
So I did philosophy, political science,
47
00:01:45.910 --> 00:01:47.440
political theory, logic,
48
00:01:47.440 --> 00:01:49.060
and I ended up doing a PhD
49
00:01:49.060 --> 00:01:50.620
in computer science at Penn.
50
00:01:50.620 --> 00:01:53.380
And you know, I sort of thought
51
00:01:53.380 --> 00:01:54.950
I was going to do testable philosophy,
52
00:01:54.950 --> 00:01:57.210
and so the closest to that was doing AI.
53
00:01:57.210 --> 00:01:58.540
So I just did this out of love.
54
00:01:58.540 --> 00:02:00.820
Like I just find the entire process
55
00:02:00.820 --> 00:02:04.550
and the goals, and the techniques, steps really fascinating.
56
00:02:04.550 --> 00:02:07.180
All part of your master plan, it all came together.
57
00:02:07.180 --> 00:02:09.880
Not at all! I just fell into it, I fell into it.
58
00:02:09.880 --> 00:02:11.820
So how did you end up then at Lyft?
59
00:02:11.820 --> 00:02:13.740
So I was at LinkedIn for about six years
60
00:02:13.740 --> 00:02:15.410
and then my wife got this phenomenal job
61
00:02:15.410 --> 00:02:17.080
at Amazon and I wanted to stay married
62
00:02:17.080 --> 00:02:19.330
so I followed her to Seattle.
63
00:02:19.330 --> 00:02:21.230
I worked for a year here at Dropbox
64
00:02:21.230 --> 00:02:22.840
and then Lyft contacted me
65
00:02:22.840 --> 00:02:25.980
and I essentially jumped at the chance because
66
00:02:25.980 --> 00:02:27.530
the space is so fascinating.
67
00:02:27.530 --> 00:02:29.990
I love cars in general
68
00:02:29.990 --> 00:02:32.210
which means I love transportation in general
69
00:02:32.210 --> 00:02:34.860
and the idea of transforming
70
00:02:34.860 --> 00:02:37.840
how we do transportation is just a fascinating space.
71
00:02:37.840 --> 00:02:39.460
And then in my prior life
72
00:02:39.460 --> 00:02:41.980
I was a tenured computer science professor,
73
00:02:41.980 --> 00:02:43.870
which is still a big love of mine,
74
00:02:43.870 --> 00:02:46.950
and so I'm an adjunct professor at Northeastern
75
00:02:46.950 --> 00:02:50.100
just to make sure I keep my teaching skills up.
76
00:02:50.100 --> 00:02:52.900
Craig, your strong humanities background
77
00:02:52.900 --> 00:02:55.200
in philosophy, political science,
78
00:02:55.200 --> 00:02:57.540
you mentioned logic, all of that.
79
00:02:57.540 --> 00:03:00.293
How did that play for you in your overall journey?
80
00:03:01.220 --> 00:03:02.580
So that's really interesting.
81
00:03:02.580 --> 00:03:04.840
When I think about what AI is,
82
00:03:04.840 --> 00:03:07.180
I find the algorithms mathematically fascinating
83
00:03:07.180 --> 00:03:08.910
but I find the use of the algorithms
84
00:03:08.910 --> 00:03:10.760
far more fascinating because
85
00:03:10.760 --> 00:03:12.450
from a technical perspective
86
00:03:12.450 --> 00:03:14.870
we're finding correlations
87
00:03:14.870 --> 00:03:17.470
in extremely high dimensional non-linear spaces.
88
00:03:17.470 --> 00:03:19.670
It's statistics at scale in some sense, right?
89
00:03:19.670 --> 00:03:22.300
We're finding these correlations between A and B
90
00:03:22.300 --> 00:03:24.030
and those algorithms are really interesting
91
00:03:24.030 --> 00:03:26.180
and I'm still teaching those now and they're fun.
92
00:03:26.180 --> 00:03:28.100
But what's more interesting to me
93
00:03:28.100 --> 00:03:30.760
is what do those correlations mean for the people?
94
00:03:30.760 --> 00:03:33.620
I think every AI model launched
95
00:03:33.620 --> 00:03:35.560
is a cognitive science test.
96
00:03:35.560 --> 00:03:38.390
Like we're trying to model the way humans behave.
97
00:03:38.390 --> 00:03:40.700
Now, for automated driving
98
00:03:40.700 --> 00:03:42.730
we're modeling the way cars behave in some sense,
99
00:03:42.730 --> 00:03:44.610
but it's really we're modeling
100
00:03:44.610 --> 00:03:47.270
the right human behavior
101
00:03:47.270 --> 00:03:49.590
given these other cars driven by humans.
102
00:03:49.590 --> 00:03:52.860
So for me, I think the goals of AI,
103
00:03:52.860 --> 00:03:55.890
I look at them much more from humanity's perspective,
104
00:03:55.890 --> 00:03:58.840
although I can nerd out on the technical side as well.
105
00:03:58.840 --> 00:03:59.780
Can you say a bit more about
106
00:03:59.780 --> 00:04:02.750
how Lyft organizes AI and ML teams?
107
00:04:02.750 --> 00:04:03.800
We have model builders
108
00:04:03.800 --> 00:04:04.780
throughout the whole company.
109
00:04:04.780 --> 00:04:06.470
We have a very large science org.
110
00:04:06.470 --> 00:04:08.330
We also have what we call ML-SWES,
111
00:04:08.330 --> 00:04:09.930
so ML software engineers.
112
00:04:09.930 --> 00:04:11.900
I run a team called Lyft ML
113
00:04:11.900 --> 00:04:13.740
and it consists of two major teams.
114
00:04:13.740 --> 00:04:15.170
One is called applied ML
115
00:04:15.170 --> 00:04:17.520
where we leverage expertise in machine learning
116
00:04:17.520 --> 00:04:19.070
to tackle some really tough problems.
117
00:04:19.070 --> 00:04:20.730
And also the ML platform,
118
00:04:20.730 --> 00:04:22.620
which drives my big interest
119
00:04:22.620 --> 00:04:24.860
in operational excellence on getting ML
120
00:04:24.860 --> 00:04:28.410
to make sure it's effectively hitting business metrics.
121
00:04:28.410 --> 00:04:30.330
What do you think, because I think Craig,
122
00:04:30.330 --> 00:04:32.010
you're still teaching, right?
123
00:04:32.010 --> 00:04:34.950
Yeah, I adjunct teach at Northeastern University
124
00:04:34.950 --> 00:04:36.160
here in Seattle.
125
00:04:36.160 --> 00:04:39.870
So, what do you think your students,
126
00:04:39.870 --> 00:04:42.280
sort of, should be asking that they're not,
127
00:04:42.280 --> 00:04:44.630
or maybe to state it in another way,
128
00:04:44.630 --> 00:04:47.510
what would they be most surprised by
129
00:04:47.510 --> 00:04:49.610
when they enter the workforce
130
00:04:49.610 --> 00:04:52.840
and actually do AI in the real world?
131
00:04:52.840 --> 00:04:55.160
The algorithms themselves
132
00:04:55.160 --> 00:04:57.780
are becoming less important.
133
00:04:57.780 --> 00:04:59.700
I'm hesitant to use the word commoditized,
134
00:04:59.700 --> 00:05:02.410
but to some degree they're being commoditized, right?
135
00:05:02.410 --> 00:05:05.630
You could pick one of five, one of seven,
136
00:05:05.630 --> 00:05:08.320
you could try them all model families
137
00:05:08.320 --> 00:05:10.090
for a particular problem.
138
00:05:10.090 --> 00:05:11.560
But what's really happening,
139
00:05:11.560 --> 00:05:14.440
or what I think is the exciting thing happening,
140
00:05:14.440 --> 00:05:16.620
is how those models fit
141
00:05:16.620 --> 00:05:19.700
into a much larger engineering pipeline
142
00:05:19.700 --> 00:05:22.620
that allow you to measure and guarantee
143
00:05:22.620 --> 00:05:25.350
that you're being effective against a business goal.
144
00:05:25.350 --> 00:05:28.250
And that has to do with the cleanliness of the data,
145
00:05:28.250 --> 00:05:30.450
making sure the data is there in a timely way,
146
00:05:30.450 --> 00:05:31.620
classic engineering things.
147
00:05:31.620 --> 00:05:35.960
Like are you returning your features at the right latency?
148
00:05:35.960 --> 00:05:38.300
So the actual model itself
149
00:05:38.300 --> 00:05:40.940
has shrunk from say, 85% of the problem
150
00:05:40.940 --> 00:05:42.540
to 15% of the problem.
151
00:05:42.540 --> 00:05:44.960
And now 85% of the problem is the engineering
152
00:05:44.960 --> 00:05:46.900
and the operational excellence surrounding it.
153
00:05:46.900 --> 00:05:49.840
And I think we're at a point of inflection there.
154
00:05:49.840 --> 00:05:52.680
So do you believe with the advent
155
00:05:52.680 --> 00:05:56.410
of auto ML and these packaged tools
156
00:05:56.410 --> 00:05:58.930
and your point about over time,
157
00:05:58.930 --> 00:06:00.740
it's less about the algo,
158
00:06:00.740 --> 00:06:03.430
more about the data and how you use it?
159
00:06:03.430 --> 00:06:06.540
Do you think the curricula, and the training,
160
00:06:06.540 --> 00:06:09.590
and just the overall orientation
161
00:06:09.590 --> 00:06:12.990
of data scientists ten years from now,
162
00:06:12.990 --> 00:06:15.110
would be dramatically different?
163
00:06:15.110 --> 00:06:16.850
Like, should we teach them different things,
164
00:06:16.850 --> 00:06:18.890
different scales? Because it used to be
165
00:06:18.890 --> 00:06:21.310
a lot is focused on creating the algorithms,
166
00:06:21.310 --> 00:06:22.920
trying different things,
167
00:06:22.920 --> 00:06:24.140
and I think you're making the point that
168
00:06:24.140 --> 00:06:27.790
that sort of plateauing, what does that mean
169
00:06:27.790 --> 00:06:31.023
in terms of the workforce of the future?
170
00:06:31.940 --> 00:06:33.660
Yeah, I think that's great.
171
00:06:33.660 --> 00:06:35.120
I'm going to say some controversial things here
172
00:06:35.120 --> 00:06:37.590
and I hope not to offend anybody.
173
00:06:37.590 --> 00:06:41.010
That's why I asked so I hope that you will!
174
00:06:41.010 --> 00:06:44.370
So if you look just five or ten years ago
175
00:06:44.370 --> 00:06:46.340
in order to deliver the kind of value
176
00:06:46.340 --> 00:06:48.530
that tech companies wanted to deliver
177
00:06:48.530 --> 00:06:52.180
you needed a fleet of PhDs, right?
178
00:06:52.180 --> 00:06:54.010
The technical ability to build those algorithms
179
00:06:54.010 --> 00:06:55.443
was extremely important.
180
00:06:56.300 --> 00:06:57.550
I think the point of inflection there
181
00:06:57.550 --> 00:07:01.200
was probably TensorFlow, 2013 ish,
182
00:07:01.200 --> 00:07:02.930
where it wasn't commoditized.
183
00:07:02.930 --> 00:07:04.220
You still needed to think very hard
184
00:07:04.220 --> 00:07:06.760
about the algorithm but the actual getting the algorithm
185
00:07:06.760 --> 00:07:08.010
out the door became a lot easier.
186
00:07:08.010 --> 00:07:10.720
Now there's plenty of frameworks for that.
187
00:07:10.720 --> 00:07:12.490
I wonder -- this is a real wonder --
188
00:07:12.490 --> 00:07:13.860
I wonder the degree to which
189
00:07:13.860 --> 00:07:15.860
we're going to need specialized machine learning,
190
00:07:15.860 --> 00:07:19.000
AI data science training going forward.
191
00:07:19.000 --> 00:07:23.220
I think CS undergrads or engineering undergrads in general
192
00:07:23.220 --> 00:07:26.810
are all going to graduate with two or three AI classes
193
00:07:26.810 --> 00:07:29.140
and those two or three AI classes
194
00:07:29.140 --> 00:07:31.330
with the right infrastructure in the company,
195
00:07:31.330 --> 00:07:32.790
the right way to gather features,
196
00:07:32.790 --> 00:07:34.950
the right way to specify your label data.
197
00:07:34.950 --> 00:07:38.460
If we have that ML platform in place,
198
00:07:38.460 --> 00:07:40.980
people with two or three strong classes
199
00:07:40.980 --> 00:07:44.030
are going to be able to deliver 70%
200
00:07:44.030 --> 00:07:45.750
of the models a company might need.
201
00:07:45.750 --> 00:07:46.810
Now that 30%,
202
00:07:46.810 --> 00:07:49.380
I think you're still going to need experts for awhile.
203
00:07:49.380 --> 00:07:50.890
I do, I just don't think you can need it
204
00:07:50.890 --> 00:07:51.740
like you used to need it,
205
00:07:51.740 --> 00:07:54.723
where almost every expert had to have a PhD.
206
00:07:55.660 --> 00:07:57.730
Yeah, I actually resonate with that Sam
207
00:07:57.730 --> 00:08:01.090
in an interesting way, it sort of corroborates
208
00:08:01.090 --> 00:08:04.240
what we've been saying about what it takes
209
00:08:04.240 --> 00:08:06.170
to actually get impact at scale
210
00:08:06.170 --> 00:08:09.830
which is like, the technical stuff gets you only so far
211
00:08:09.830 --> 00:08:12.830
but ultimately you have to change the way it's consumed,
212
00:08:12.830 --> 00:08:15.040
and you have to change the way people work,
213
00:08:15.040 --> 00:08:18.703
and the different modes of interaction between humans and
214
00:08:19.710 --> 00:08:21.440
AI, I guess that's a lot of the humanities,
215
00:08:21.440 --> 00:08:23.410
and the philosophy, and the political science,
216
00:08:23.410 --> 00:08:25.710
and the how, sort of, the human works
217
00:08:25.710 --> 00:08:28.600
more so than what the algo does.
218
00:08:28.600 --> 00:08:29.960
Well that's a good redirection too,
219
00:08:29.960 --> 00:08:31.040
because if we're not careful,
220
00:08:31.040 --> 00:08:32.530
then that conversation slips us
221
00:08:32.530 --> 00:08:36.430
into the curriculum being DevOps more.
222
00:08:36.430 --> 00:08:38.720
And so what Shervin is pointing out is that
223
00:08:38.720 --> 00:08:40.600
maybe that's a component, of course too,
224
00:08:40.600 --> 00:08:42.210
but there's process change
225
00:08:42.210 --> 00:08:45.693
and more, let's say business-oriented initiatives.
226
00:08:46.710 --> 00:08:47.820
So what other kinds of things
227
00:08:47.820 --> 00:08:48.920
are you trying to teach people?
228
00:08:48.920 --> 00:08:49.940
Or what other kinds of things
229
00:08:49.940 --> 00:08:51.640
do you think executives should know?
230
00:08:51.640 --> 00:08:53.330
I mean, we can't have the ...
231
00:08:53.330 --> 00:08:55.120
everybody can't have to know everything.
232
00:08:55.120 --> 00:08:56.890
Then it would be a bit overwhelming.
233
00:08:56.890 --> 00:08:58.250
I mean that perhaps that's ideal
234
00:08:58.250 --> 00:08:59.730
if everyone knows everything,
235
00:08:59.730 --> 00:09:02.490
but what exactly do different levels
236
00:09:02.490 --> 00:09:04.003
of managers need to know?
237
00:09:05.170 --> 00:09:06.930
I think the top decision maker
238
00:09:06.930 --> 00:09:11.100
needs to understand dangers of a model going awry
239
00:09:11.100 --> 00:09:14.450
and they need to understand the overall process
240
00:09:14.450 --> 00:09:16.690
that you really need label data.
241
00:09:16.690 --> 00:09:17.830
Like there's no magic here.
242
00:09:17.830 --> 00:09:20.475
That they have to understand there's not magic, right?
243
00:09:20.475 --> 00:09:22.050
So they have to understand
244
00:09:22.050 --> 00:09:24.320
that label data is expensive,
245
00:09:24.320 --> 00:09:26.300
that getting the labels right
246
00:09:26.300 --> 00:09:28.030
and sampling the distribution
247
00:09:28.030 --> 00:09:29.460
of the world that you want correctly
248
00:09:29.460 --> 00:09:31.390
is extremely important.
249
00:09:31.390 --> 00:09:32.960
I believe they also have to understand
250
00:09:32.960 --> 00:09:33.990
the life cycle in general,
251
00:09:33.990 --> 00:09:36.250
which is different than, you know, two week sprints.
252
00:09:36.250 --> 00:09:38.980
We're going to close these JIRA tickets, right?
253
00:09:38.980 --> 00:09:40.817
That data gathering is extremely important
254
00:09:40.817 --> 00:09:42.600
and that could take a quarter or two.
255
00:09:42.600 --> 00:09:44.920
And that the first model you ship
256
00:09:44.920 --> 00:09:47.200
probably isn't going to be very good,
257
00:09:47.200 --> 00:09:49.910
you know, because it was from a small label dataset
258
00:09:49.910 --> 00:09:51.770
and now you're gathering data in the wild.
259
00:09:51.770 --> 00:09:53.590
So there's a life cycle piece
260
00:09:53.590 --> 00:09:55.090
that they need to understand.
261
00:09:55.090 --> 00:09:57.370
And they need to understand that,
262
00:09:57.370 --> 00:09:59.410
unfortunately in a lot of ways,
263
00:09:59.410 --> 00:10:02.380
maybe not for car driving but for recommendations,
264
00:10:02.380 --> 00:10:04.840
the first couple that you ship get iteratively better.
265
00:10:04.840 --> 00:10:06.770
So I think that's extremely important for the top.
266
00:10:06.770 --> 00:10:08.610
I think for a couple of levels down,
267
00:10:08.610 --> 00:10:09.500
they need to understand
268
00:10:09.500 --> 00:10:11.160
like the precision-recall trade-off,
269
00:10:11.160 --> 00:10:13.920
the kinds of errors your model can make.
270
00:10:13.920 --> 00:10:16.070
Your model can either be making false negative errors
271
00:10:16.070 --> 00:10:17.330
or false positive errors
272
00:10:17.330 --> 00:10:20.160
and I think it's extremely important as a product person
273
00:10:20.160 --> 00:10:22.260
that you own that choice.
274
00:10:22.260 --> 00:10:24.110
So if we're doing document search,
275
00:10:24.110 --> 00:10:26.230
I think you care a lot more about false positives.
276
00:10:26.230 --> 00:10:27.500
You care a lot more about precision.
277
00:10:27.500 --> 00:10:31.010
You want the things that come to the top to be relevant.
278
00:10:31.010 --> 00:10:33.180
And for most search problems,
279
00:10:33.180 --> 00:10:34.880
you don't have to get all the relevant things.
280
00:10:34.880 --> 00:10:36.570
You just have to get enough of the relevant things.
281
00:10:36.570 --> 00:10:38.700
So if some relevant things are called non-relevant
282
00:10:38.700 --> 00:10:40.570
you're okay with that, right?
283
00:10:40.570 --> 00:10:42.030
But for other problems,
284
00:10:42.030 --> 00:10:43.930
you need to get everything.
285
00:10:43.930 --> 00:10:45.130
Document search, that's fine.
286
00:10:45.130 --> 00:10:47.320
But yeah, Lyft as well, like put it in the context
287
00:10:47.320 --> 00:10:50.080
of one of these companies where you've had a precision
288
00:10:50.080 --> 00:10:53.220
and recall trade-off, false positive, false negative.
289
00:10:53.220 --> 00:10:54.930
I think luckily at Lyft
290
00:10:54.930 --> 00:10:56.890
we have nice human escape hatches
291
00:10:56.890 --> 00:10:58.140
which I think is extremely important.
292
00:10:58.140 --> 00:10:59.780
Like all these recommendations
293
00:10:59.780 --> 00:11:01.580
ideally should have a human escape hatch.
294
00:11:01.580 --> 00:11:03.910
So if I recommend a destination for you
295
00:11:03.910 --> 00:11:05.410
and that destination is wrong,
296
00:11:06.630 --> 00:11:09.073
no harm, no foul, you just type the destination in.
297
00:11:10.060 --> 00:11:12.020
So I think for Lyft as a product
298
00:11:12.020 --> 00:11:13.260
I think we're pretty lucky
299
00:11:13.260 --> 00:11:15.330
because most of our recommendations,
300
00:11:15.330 --> 00:11:16.580
which are trying to lower friction
301
00:11:16.580 --> 00:11:17.940
to get you to take a ride,
302
00:11:17.940 --> 00:11:20.690
it's okay if we don't get them exactly right.
303
00:11:20.690 --> 00:11:22.000
There's no real danger there.
304
00:11:22.000 --> 00:11:24.070
Self-driving cars, that one's tough
305
00:11:24.070 --> 00:11:25.530
because you want to get them both.
306
00:11:25.530 --> 00:11:27.440
You want to know that's a pedestrian and you
307
00:11:27.440 --> 00:11:30.040
also want to make sure you don't miss any pedestrians.
308
00:11:30.040 --> 00:11:31.960
And the idea of putting a human in the loop
309
00:11:31.960 --> 00:11:33.603
there is much more problematic than just saying,
310
00:11:33.603 --> 00:11:35.420
all right, here's some destinations,
311
00:11:35.420 --> 00:11:36.520
which one do you like?
312
00:11:37.960 --> 00:11:40.840
Craig, earlier you talked about,
313
00:11:40.840 --> 00:11:43.160
you know, how AI in real life
314
00:11:43.160 --> 00:11:45.850
is like a bunch of cognitive science experiments
315
00:11:45.850 --> 00:11:47.300
because it's ultimately about-
316
00:11:47.300 --> 00:11:48.380
For me at least.
317
00:11:48.380 --> 00:11:52.008
Yeah, and it brought up the idea of
318
00:11:52.008 --> 00:11:56.150
unconscious bias. And so like we as humans have become
319
00:11:56.150 --> 00:11:59.766
a lot more aware about our unconscious biases
320
00:11:59.766 --> 00:12:01.550
across everything, right?
321
00:12:01.550 --> 00:12:03.380
Because they've been ingrained
322
00:12:03.380 --> 00:12:06.910
through generations and stereotypes, et cetera.
323
00:12:06.910 --> 00:12:08.710
And just our past experience, right?
324
00:12:08.710 --> 00:12:11.670
Like a biased world creates a biased experience
325
00:12:11.670 --> 00:12:14.290
even if you have the best possible intentions.
326
00:12:14.290 --> 00:12:15.610
Exactly, right?
327
00:12:15.610 --> 00:12:19.450
And so I guess my question is,
328
00:12:20.340 --> 00:12:24.763
clearly there is unintended bias in AI, has to be.
329
00:12:26.650 --> 00:12:29.100
What do you think we need to think about now
330
00:12:29.100 --> 00:12:31.920
so that 10, 20 years from now
331
00:12:32.870 --> 00:12:36.100
that bias hasn't become so ingrained
332
00:12:36.100 --> 00:12:39.130
in how AI works that it would be so hard
333
00:12:39.130 --> 00:12:40.643
to then course correct?
334
00:12:42.230 --> 00:12:43.063
(laughs) It already has.
335
00:12:43.063 --> 00:12:45.200
So the question is how do we course correct?
336
00:12:45.200 --> 00:12:48.530
So let me start by saying I was on a panel
337
00:12:48.530 --> 00:12:51.880
for Northeastern about this movie "Coded Bias."
338
00:12:51.880 --> 00:12:53.600
So if you haven't seen the movie "Coded Bias"
339
00:12:53.600 --> 00:12:55.100
you should absolutely see it.
340
00:12:55.100 --> 00:12:58.770
It's about this MIT media lab undergraduate black woman
341
00:12:58.770 --> 00:13:01.240
who tried to do a project
342
00:13:01.240 --> 00:13:03.720
that didn't work because facial recognition
343
00:13:03.720 --> 00:13:06.540
just simply didn't work for black females.
344
00:13:06.540 --> 00:13:10.070
It's just an absolutely fascinating social study.
345
00:13:10.070 --> 00:13:12.860
The dataset that was used
346
00:13:12.860 --> 00:13:16.130
to train the machine learning,
347
00:13:16.130 --> 00:13:18.830
so the facial recognition algorithm,
348
00:13:18.830 --> 00:13:22.440
was gathered by the researchers at the time,
349
00:13:22.440 --> 00:13:23.740
and the researchers at the time
350
00:13:23.740 --> 00:13:25.410
were a bunch of white males.
351
00:13:25.410 --> 00:13:26.950
And this is a known issue, right?
352
00:13:26.950 --> 00:13:28.870
There's a skew in the way the dataset is gathered.
353
00:13:28.870 --> 00:13:31.730
Look, there's a similar skew in all psychological studies.
354
00:13:31.730 --> 00:13:33.500
Psychological studies don't apply to me.
355
00:13:33.500 --> 00:13:36.410
I'm 56, psychological studies apply to college students
356
00:13:36.410 --> 00:13:39.450
because that's the readily available subjects, right?
357
00:13:39.450 --> 00:13:41.580
So these were the readily available people
358
00:13:41.580 --> 00:13:43.240
because of the biased world
359
00:13:43.240 --> 00:13:45.090
and so that's how the dataset came about.
360
00:13:45.090 --> 00:13:46.750
So even if no ill intention
361
00:13:46.750 --> 00:13:48.170
the world was skewed,
362
00:13:48.170 --> 00:13:49.060
the world was biased,
363
00:13:49.060 --> 00:13:50.070
data was biased,
364
00:13:50.070 --> 00:13:53.000
it didn't work for a great number of people.
365
00:13:53.000 --> 00:13:55.530
And not a lot of females were part of the training set
366
00:13:55.530 --> 00:13:58.760
and then the darker your skin, the worse it got.
367
00:13:58.760 --> 00:14:00.220
And there's all kinds of technical reasons
368
00:14:00.220 --> 00:14:02.890
why darker skin has less contrast, blah, blah, blah,
369
00:14:02.890 --> 00:14:06.410
but that's not the issue.
370
00:14:06.410 --> 00:14:10.660
The issue is, should we have gathered the data that way?
371
00:14:10.660 --> 00:14:12.860
What is the goal of the dataset?
372
00:14:12.860 --> 00:14:14.150
Who are our customers?
373
00:14:14.150 --> 00:14:15.530
Who do we want to serve?
374
00:14:15.530 --> 00:14:17.200
And let's sample the data
375
00:14:17.200 --> 00:14:19.870
in such a way that it's serving our customers.
376
00:14:19.870 --> 00:14:21.480
We talked about this earlier about the undergrads,
377
00:14:21.480 --> 00:14:22.860
I think that's really important.
378
00:14:22.860 --> 00:14:24.570
One way to get out of that
379
00:14:24.570 --> 00:14:26.330
is diversity in the workplace.
380
00:14:26.330 --> 00:14:27.740
I believe this so strongly.
381
00:14:27.740 --> 00:14:31.830
And you ask everybody, all of these diverse groups,
382
00:14:31.830 --> 00:14:33.440
to test the system
383
00:14:33.440 --> 00:14:36.120
and to see if the system works for them.
384
00:14:36.120 --> 00:14:38.240
When we did image search at Dropbox
385
00:14:38.240 --> 00:14:40.350
we asked all of the employee research groups,
386
00:14:40.350 --> 00:14:41.340
please search for things
387
00:14:41.340 --> 00:14:43.890
that in the past have been problematic for you
388
00:14:43.890 --> 00:14:45.380
and see if we got them right.
389
00:14:45.380 --> 00:14:47.440
And if we found some that were wrong,
390
00:14:47.440 --> 00:14:50.020
we would go back and regather data
391
00:14:50.020 --> 00:14:51.970
to mitigate against those issues.
392
00:14:51.970 --> 00:14:54.160
So look, your system is going to be biased
393
00:14:54.160 --> 00:14:56.190
by the data that's gathered, fact, just a fact.
394
00:14:56.190 --> 00:14:57.740
It's going to be biased by the data that's gathered.
395
00:14:57.740 --> 00:14:59.770
You want to do your best to gather it correctly.
396
00:14:59.770 --> 00:15:01.420
You're probably not going to gather it correctly
397
00:15:01.420 --> 00:15:03.330
because you have your own unconscious bias
398
00:15:03.330 --> 00:15:04.163
as you point out.
399
00:15:04.163 --> 00:15:05.440
So you have to ask all the people
400
00:15:05.440 --> 00:15:07.090
who are going to be your customers
401
00:15:07.090 --> 00:15:09.060
to try it, to bang on it,
402
00:15:09.060 --> 00:15:11.090
to make sure it's doing the right thing.
403
00:15:11.090 --> 00:15:12.940
And when it's not, go back and gather
404
00:15:12.940 --> 00:15:14.320
the data necessary to fix it.
405
00:15:14.320 --> 00:15:18.100
So I think the short answer is diversity in the workplace.
406
00:15:18.100 --> 00:15:19.310
Craig, thanks for taking the time
407
00:15:19.310 --> 00:15:20.220
to talk with us today.
408
00:15:20.220 --> 00:15:22.120
Lots of interesting things.
409
00:15:22.120 --> 00:15:24.360
My pleasure, these are really fun conversations.
410
00:15:24.360 --> 00:15:27.810
I'm pretty nerdy about this so I enjoyed it very much.
411
00:15:27.810 --> 00:15:29.390
Your enthusiasm shows.
412
00:15:29.390 --> 00:15:30.530
Really insightful stuff,
413
00:15:30.530 --> 00:15:31.893
thank you.
Thank you guys.
414
00:15:35.670 --> 00:15:36.930
Well Shervin, Craig says
415
00:15:36.930 --> 00:15:38.610
he won the lottery in his career
416
00:15:38.610 --> 00:15:39.670
but I think we won the lottery
417
00:15:39.670 --> 00:15:40.680
in getting him as a guest
418
00:15:40.680 --> 00:15:42.860
for our first episode of season two.
419
00:15:42.860 --> 00:15:43.703
Let's recap.
420
00:15:44.590 --> 00:15:47.050
I mean, he made a lot of good points.
421
00:15:47.050 --> 00:15:51.160
Clearly the commoditization of algorithms over time,
422
00:15:51.160 --> 00:15:54.750
and how it's more and more going to be about
423
00:15:54.750 --> 00:15:58.040
tying it with strategy going back to key business metrics,
424
00:15:58.040 --> 00:16:01.330
making change happen, the usage.
425
00:16:01.330 --> 00:16:03.040
I really liked his point on
426
00:16:03.040 --> 00:16:07.630
what it takes to get the bias out of the system
427
00:16:07.630 --> 00:16:10.890
and how bias is already in the system.
428
00:16:10.890 --> 00:16:12.450
The commoditization is particularly important.
429
00:16:12.450 --> 00:16:13.500
I think it resonates with us
430
00:16:13.500 --> 00:16:14.720
because we were talking about this
431
00:16:14.720 --> 00:16:15.910
from a business perspective.
432
00:16:15.910 --> 00:16:17.430
And so what he's saying is that,
433
00:16:17.430 --> 00:16:18.780
a lot of this is going to become
434
00:16:18.780 --> 00:16:21.080
increasingly a business problem.
435
00:16:21.080 --> 00:16:23.570
When it's a business problem it's not a technical problem.
436
00:16:23.570 --> 00:16:25.590
I don't want to discount the technical aspects of it
437
00:16:25.590 --> 00:16:26.580
and certainly, you know,
438
00:16:26.580 --> 00:16:29.290
he brings plenty of technical chops to the table,
439
00:16:29.290 --> 00:16:31.530
but he really reinforced the
440
00:16:31.530 --> 00:16:33.900
this is a business problem now aspect.
441
00:16:33.900 --> 00:16:35.590
Yeah, I mean in five minutes
442
00:16:35.590 --> 00:16:38.620
he basically provided such a cogent argument
443
00:16:38.620 --> 00:16:41.700
for our last two reports, right?
444
00:16:41.700 --> 00:16:44.500
The 2019, 2020, it's about strategy
445
00:16:44.500 --> 00:16:47.760
and process change and process redesign and reengineering
446
00:16:47.760 --> 00:16:52.200
and it's about human and AI interaction and adoption.
447
00:16:52.200 --> 00:16:53.910
And what's also a business problem too
448
00:16:53.910 --> 00:16:55.430
is the managerial choice.
449
00:16:55.430 --> 00:16:57.120
I mean, he came back to that as well.
450
00:16:57.120 --> 00:16:59.230
He was talking about some of these things
451
00:16:59.230 --> 00:17:01.290
are not clear-cut decisions.
452
00:17:01.290 --> 00:17:04.350
There's a choice between which way you make a mistake.
453
00:17:04.350 --> 00:17:07.290
That's a management problem not a technical problem.
454
00:17:07.290 --> 00:17:10.520
And it also requires managers
455
00:17:10.520 --> 00:17:11.730
to know what they're talking about,
456
00:17:11.730 --> 00:17:14.080
which means they need to really, really understand
457
00:17:14.080 --> 00:17:17.530
what AI is saying and what it could be saying
458
00:17:17.530 --> 00:17:19.840
and what's its limitations,
459
00:17:19.840 --> 00:17:21.750
and what's the art of the possible.
460
00:17:21.750 --> 00:17:23.690
And I also really liked the point that
461
00:17:23.690 --> 00:17:26.690
as you get closer to the developers
462
00:17:26.690 --> 00:17:28.280
and the builders of AI,
463
00:17:28.280 --> 00:17:30.170
you have to really, really understand the math
464
00:17:30.170 --> 00:17:33.120
and the code because otherwise you can't guide them.
465
00:17:33.120 --> 00:17:33.980
Although don't you worry
466
00:17:33.980 --> 00:17:35.280
that we're just running into this thing
467
00:17:35.280 --> 00:17:37.040
where everyone has to understand everything?
468
00:17:37.040 --> 00:17:38.790
I feel like that's a tough sell.
469
00:17:38.790 --> 00:17:40.200
Like if the managers have to understand
470
00:17:40.200 --> 00:17:41.460
the business and how to make money
471
00:17:41.460 --> 00:17:43.020
and they have to understand the code.
472
00:17:43.020 --> 00:17:46.060
I mean, having everyone understand everything is obviously-
473
00:17:46.060 --> 00:17:47.210
Well I guess the question is
474
00:17:47.210 --> 00:17:49.810
how much do you have to understand everything?
475
00:17:49.810 --> 00:17:53.050
I mean, a good business executive
476
00:17:53.050 --> 00:17:54.540
already understands everything
477
00:17:54.540 --> 00:17:57.000
to the level that he or she should,
478
00:17:57.000 --> 00:17:59.760
to the point of asking the right questions.
479
00:17:59.760 --> 00:18:01.400
I think you're right.
480
00:18:01.400 --> 00:18:04.410
But I think, isn't this like what Einstein said that
481
00:18:05.310 --> 00:18:07.190
you don't really understand something
482
00:18:07.190 --> 00:18:10.330
unless you can describe it to a five year old.
483
00:18:10.330 --> 00:18:13.340
You know, you can describe gravity to a five year old
484
00:18:13.340 --> 00:18:16.380
and to a 20 year old and to a grad student
485
00:18:16.380 --> 00:18:19.290
in different ways and they will all understand it.
486
00:18:19.290 --> 00:18:21.270
The question is, at least you understand it
487
00:18:21.270 --> 00:18:22.830
rather than just say, I have no idea
488
00:18:22.830 --> 00:18:24.390
there is such a thing as gravity.
489
00:18:24.390 --> 00:18:26.440
So basically teaching and academics
490
00:18:26.440 --> 00:18:27.273
are really important.
491
00:18:27.273 --> 00:18:30.020
Is that what Shervin has just gone on the record as saying?
492
00:18:30.020 --> 00:18:33.070
I think the idea that managers
493
00:18:33.070 --> 00:18:36.193
and senior executives need to understand AI,
494
00:18:37.470 --> 00:18:39.320
itself is not a slam dunk
495
00:18:39.320 --> 00:18:40.900
because you're raising the right question.
496
00:18:40.900 --> 00:18:42.640
What is the right level of understanding?
497
00:18:42.640 --> 00:18:44.810
And so what is the right level of synthesis
498
00:18:44.810 --> 00:18:47.650
and articulation that allows you
499
00:18:47.650 --> 00:18:48.870
to make the right decisions
500
00:18:48.870 --> 00:18:51.100
without having to know everything,
501
00:18:51.100 --> 00:18:54.550
but isn't that what a successful business executive
502
00:18:54.550 --> 00:18:57.840
does with every business problem?
503
00:18:57.840 --> 00:18:59.590
And I think that's what we're saying,
504
00:18:59.590 --> 00:19:03.610
that with AI, you need to know enough to be able to probe
505
00:19:03.610 --> 00:19:06.400
but suffice it to say, it's not a black box.
506
00:19:06.400 --> 00:19:09.050
Like a lot of the technology implementations
507
00:19:09.050 --> 00:19:11.170
have been a black box in the past.
508
00:19:11.170 --> 00:19:12.110
And that helps get back
509
00:19:12.110 --> 00:19:13.830
to the whole learning more,
510
00:19:13.830 --> 00:19:15.860
and where to draw the line,
511
00:19:15.860 --> 00:19:18.140
and help to understand that balance.
512
00:19:18.140 --> 00:19:20.150
I guess after the discussion of gravity
513
00:19:20.150 --> 00:19:21.600
each one of those people would understand
514
00:19:21.600 --> 00:19:23.610
more about gravity than they did before.
515
00:19:23.610 --> 00:19:26.320
And so it's a matter of moving from current state
516
00:19:26.320 --> 00:19:27.709
to next state.
517
00:19:27.709 --> 00:19:28.820
Yeah.
518
00:19:28.820 --> 00:19:30.140
Craig made some important points
519
00:19:30.140 --> 00:19:31.820
about diversity in the workplace.
520
00:19:31.820 --> 00:19:33.280
If the team gathering data
521
00:19:33.280 --> 00:19:35.300
isn't hyper aware of the inherent biases
522
00:19:35.300 --> 00:19:36.720
in their data sets,
523
00:19:36.720 --> 00:19:39.313
algorithms are destined to produce a biased result.
524
00:19:40.150 --> 00:19:42.260
He refers to the movie "Coded Bias"
525
00:19:42.260 --> 00:19:45.113
and the MIT media lab researcher, Joy Buolamwini.
526
00:19:45.970 --> 00:19:49.200
Joy is the founder of the Algorithmic Justice League.
527
00:19:49.200 --> 00:19:50.850
We'll provide some links in the show notes
528
00:19:50.850 --> 00:19:53.720
where you can read more about Joy and her research.
529
00:19:53.720 --> 00:19:55.070
Thanks for joining us today.
530
00:19:55.070 --> 00:19:56.650
We're looking forward to the next episode
531
00:19:56.650 --> 00:19:58.200
when we'll talk with Will Grannis
532
00:19:58.200 --> 00:19:59.640
who has a unique challenge of building
533
00:19:59.640 --> 00:20:01.763
the CTO function at Google Cloud.
534
00:20:02.720 --> 00:20:03.673
Until next time.
535
00:20:05.990 --> 00:20:08.730
Thanks for listening to "Me, Myself, and AI".
536
00:20:08.730 --> 00:20:10.340
If you're enjoying this show
537
00:20:10.340 --> 00:20:12.450
take a minute to write us a review.
538
00:20:12.450 --> 00:20:14.080
If you send us a screenshot
539
00:20:14.080 --> 00:20:17.370
we'll send you a collection of MIT SMR's best articles
540
00:20:17.370 --> 00:20:20.910
on artificial intelligence free for a limited time.
541
00:20:20.910 --> 00:20:22.380
Send your review screenshot
542
00:20:22.380 --> 00:20:25.776
to smrfeedback@mit.edu.
543
00:20:25.776 --> 00:20:29.609
(soft upbeat music continues)