back to index

Michael Kearns: Algorithmic Fairness, Privacy & Ethics | Lex Fridman Podcast #50


small model | large model

link |
00:00:00.000
The following is a conversation with Michael Kearns. He's a professor at the University of
link |
00:00:05.040
Pennsylvania and a coauthor of the new book, Ethical Algorithm, that is the focus of much of
link |
00:00:10.880
this conversation. It includes algorithmic fairness, bias, privacy, and ethics in general.
link |
00:00:18.080
But that is just one of many fields that Michael is a world class researcher in,
link |
00:00:22.480
some of which we touch on quickly, including learning theory or the theoretical foundation
link |
00:00:27.840
of machine learning, game theory, quantitative finance, computational social science, and much
link |
00:00:33.120
more. But on a personal note, when I was an undergrad early on, I worked with Michael on
link |
00:00:39.680
an algorithmic trading project and competition that he led. That's when I first fell in love
link |
00:00:44.720
with algorithmic game theory. While most of my research life has been in machine learning and
link |
00:00:49.600
human robot interaction, the systematic way that game theory reveals the beautiful structure in our
link |
00:00:55.440
competitive and cooperating world of humans has been a continued inspiration to me. So for that
link |
00:01:02.960
and other things, I'm deeply thankful to Michael and really enjoyed having this conversation again
link |
00:01:08.960
in person after so many years. This is the Artificial Intelligence Podcast. If you enjoy it,
link |
00:01:15.280
subscribe on YouTube, give it five stars on Apple Podcasts, support on Patreon,
link |
00:01:20.160
or simply connect with me on Twitter. Alex Friedman spelled F R I D M A N. This episode
link |
00:01:27.120
is supported by an amazing podcast called pessimists archive. Jason, the host of the show,
link |
00:01:33.200
reached out to me looking to support this podcast. And so I listened to it to check it out. And by
link |
00:01:39.040
listened, I mean, I went through it Netflix binge style, at least five episodes in a row.
link |
00:01:45.280
It's not one of my favorite podcasts. And I think it should be one of the top podcasts in
link |
00:01:49.280
the world, frankly. It's a history show about why people resist new things. Each episode looks at
link |
00:01:55.760
a moment in history, when something new was introduced, something that today we think of
link |
00:02:00.560
as commonplace, like recorded music, umbrellas, bicycles, cars, chess, coffee, the elevator,
link |
00:02:06.880
and the show explores why it freaked everyone out. The latest episode on mirrors and vanity
link |
00:02:12.720
still stays with me as I think about vanity in the modern day of the Twitter world.
link |
00:02:17.600
That's the fascinating thing about the show is that stuff that happened long ago, especially in
link |
00:02:22.000
terms of our fear of new things, repeats itself in the modern day. And so it has many lessons for
link |
00:02:26.960
us to think about in terms of human psychology, and the role of technology in our society.
link |
00:02:32.880
Anyway, you should subscribe and listen to pessimist archive. I highly recommend it. And now here's my
link |
00:02:39.840
conversation with Michael Kearns. You mentioned reading Fear and Loading in Las Vegas in high
link |
00:02:48.160
school and having more or a bit more of a literary mind. So what books, non technical,
link |
00:02:55.600
non computer science, would you say had the biggest impact on your life, either intellectually or
link |
00:03:02.000
emotionally? You've dug deep into my history, I see. When deep? Yeah, I think my favorite
link |
00:03:07.760
novel is Infinite Jest by David Foster Wallace, which actually coincidentally, much of it takes
link |
00:03:14.800
place in the halls of buildings right around us here at MIT. So that certainly had a big influence
link |
00:03:20.160
on me. And as you noticed, like when I was in high school, I actually even started college
link |
00:03:25.440
as an English major. So was very influenced by sort of that genre of journalism at the time and
link |
00:03:31.040
thought I wanted to be a writer and then realized that an English major teaches you to read, but
link |
00:03:35.440
it doesn't teach you how to write. And then I became interested in math and computer science
link |
00:03:39.280
instead. Well, in your new book, Ethical Algorithm, you kind of sneak up from an algorithmic
link |
00:03:46.720
perspective on these deep profound philosophical questions of fairness, of privacy. In thinking
link |
00:03:56.960
about these topics, how often do you return to that literary mind that you had? Yeah, I'd like to
link |
00:04:03.840
claim there was a deeper connection. But I think both Aaron and I kind of came at these topics
link |
00:04:10.720
first and foremost from a technical angle. I mean, I kind of consider myself primarily
link |
00:04:17.040
and originally a machine learning researcher. And I think as we just watched like the rest
link |
00:04:21.760
of the society, the field technically advance, and then quickly on the heels of that kind of the
link |
00:04:26.800
buzzkill of all of the antisocial behavior by algorithms, just kind of realized there was
link |
00:04:31.680
an opportunity for us to do something about it from a research perspective.
link |
00:04:37.200
More to the point in your question, I mean, I do have an uncle who is literally a moral
link |
00:04:42.480
philosopher. And so in the early days of our technical work on fairness topics, I would
link |
00:04:48.000
occasionally run ideas behind him. So I mean, I remember an early email I sent to him in which
link |
00:04:52.640
I said like, oh, here's a specific definition of algorithmic fairness that we think is some sort
link |
00:04:58.480
of variant of Rawlsian fairness. What do you think? And I thought I was asking a yes or no
link |
00:05:04.960
question. And I got back here, kind of classical philosophers, responding, well, it depends.
link |
00:05:09.760
If you look at it this way, then you might conclude this. And that's when I realized that there was
link |
00:05:14.960
a real kind of rift between the ways philosophers and others had thought about things like fairness,
link |
00:05:23.120
you know, from sort of a humanitarian perspective, and the way that you needed to think about it
link |
00:05:27.840
as a computer scientist, if you were going to kind of implement actual algorithmic solutions.
link |
00:05:34.080
But I would say the algorithmic solutions take care of some of the low hanging fruit,
link |
00:05:41.360
sort of the problem is a lot of algorithms, when they don't consider fairness,
link |
00:05:47.280
they are just terribly unfair. And when they don't consider privacy, they're terribly,
link |
00:05:52.320
they violate privacy, sort of the algorithmic approach fixes big problems. But there's still,
link |
00:06:01.520
you get, when you start pushing into the gray area, that's when you start getting to this
link |
00:06:05.360
philosophy of what it means to be fair, starting from Plato, what is justice kind of questions?
link |
00:06:12.400
Yeah, I think that's right. And I mean, I would even not go as far as you went to say that sort
link |
00:06:17.920
of the algorithmic work in these areas is solving like the biggest problems. And, you know, we
link |
00:06:23.520
discussed in the book, the fact that really we are, there's a sense in which we're kind of looking
link |
00:06:28.400
where the light is in that, you know, for example, if police are racist in who they decide to stop
link |
00:06:36.480
and frisk, and that goes into the data, there's sort of no undoing that downstream by kind of
link |
00:06:42.080
clever algorithmic methods. And I think especially in fairness, I mean, I think less so in privacy
link |
00:06:49.760
where we feel like the community kind of really has settled on the right definition, which is
link |
00:06:54.720
differential privacy. If you just look at the algorithmic fairness literature already, you
link |
00:06:59.360
can see it's going to be much more of a mess. And you know, you've got these theorems saying,
link |
00:07:03.920
here are three entirely reasonable, desirable notions of fairness. And you know, here's a proof
link |
00:07:11.600
that you cannot simultaneously have all three of them. So I think we know that algorithmic
link |
00:07:17.600
fairness compared to algorithmic privacy is going to be kind of a harder problem.
link |
00:07:22.080
And it will have to revisit, I think, things that have been thought about
link |
00:07:26.160
by, you know, many generations of scholars before us. So it's very early days for fairness, I think.
link |
00:07:32.800
So before we get into the details of differential privacy and then the fairness side,
link |
00:07:37.280
I mean, linger on the philosophy, but do you think most people are fundamentally good?
link |
00:07:43.600
Or do most of us have both the capacity for good and evil within us?
link |
00:07:48.640
I mean, I'm an optimist. I tend to think that most people are good and want to do to do right.
link |
00:07:55.520
And that deviations from that are, you know, kind of usually due to circumstance, not due to people
link |
00:08:01.840
being bad at heart. With people with power are people at the heads of governments,
link |
00:08:08.160
people at the heads of companies, people at the heads of maybe so financial power markets.
link |
00:08:14.960
Do you think the distribution there is also most people are good and have good intent?
link |
00:08:21.920
Yeah, I do. I mean, my statement wasn't qualified to people not in positions of power. I mean,
link |
00:08:28.800
I think what happens in a lot of the, you know, the cliche about absolute power corrupts absolutely.
link |
00:08:34.080
I mean, you know, I think even short of that, you know, having spent a lot of time on Wall Street
link |
00:08:40.720
and also in arenas very, very different from Wall Street like academia, you know, one of the things
link |
00:08:47.920
I think I benefited from by moving between two very different worlds is you become aware that,
link |
00:08:54.000
you know, these worlds kind of develop their own social norms and they develop their own
link |
00:08:59.440
rationales for, you know, behavior, for instance, that might look unusual to outsiders. But when
link |
00:09:05.600
you're in that world, it doesn't feel unusual at all. And I think this is true of a lot of,
link |
00:09:11.680
you know, professional cultures, for instance. And, you know, so then your maybe slippery slope
link |
00:09:18.480
is too strong of a word, but, you know, you're in some world where you're mainly around other people
link |
00:09:23.200
with the same kind of viewpoints and training and worldview as you. And I think that's more of a source
link |
00:09:30.080
of, you know, kind of abuses of power than sort of, you know, there being good people and evil
link |
00:09:37.200
people and that somehow the evil people are the ones that somehow rise to power.
link |
00:09:42.560
That's really interesting. So it's the, within the social norms constructed by that particular
link |
00:09:48.240
group of people, you're all trying to do good. But because it's a group, you might be, you might
link |
00:09:54.320
drift into something that for the broader population, it does not align with the values of
link |
00:09:59.280
society. That kind of, that's the word. Yeah. I mean, or not that you drift, but even that
link |
00:10:03.920
things that don't make sense to the outside world don't seem unusual to you. So it's not
link |
00:10:11.200
sort of like a good or a bad thing. But, you know, like so, for instance, you know, on in the world
link |
00:10:16.400
in the world of finance, right, there's a lot of complicated types of activity that if you are not
link |
00:10:22.000
immersed in that world, you cannot see why the purpose of that, you know, that activity exists
link |
00:10:27.040
at all. It just seems like, you know, completely useless and people just like, you know, pushing
link |
00:10:32.160
money around. And when you're in that world, right, you're, you're, and you learn more, your
link |
00:10:37.040
view does become more nuanced, right? You realize, okay, there is actually a function to this activity.
link |
00:10:43.040
And for in some cases, you would conclude that actually, if magically we could eradicate this
link |
00:10:47.920
activity tomorrow, it would come back because it actually is like serving some useful purpose.
link |
00:10:53.600
It's just a useful purpose is very difficult for outsiders to see. And so I think, you know, lots of
link |
00:11:00.000
professional work environments or cultures, as I might put it, kind of have these social norms
link |
00:11:07.600
that, you know, don't make sense to the outside world, academia is the same, right? I mean,
link |
00:11:11.040
lots of people look at academia and say, you know, what the hell are all of you people doing?
link |
00:11:15.840
Yeah. Why are you paid so much? In some cases, a taxpayer expenses to do, you know, to, you know,
link |
00:11:22.480
publish papers that nobody reads, you know, but when you're in that world, you come to see the
link |
00:11:26.720
value for it. And but even though you might not be able to explain it to, you know, the person in
link |
00:11:31.360
the street. Right. And in the case of the financial sector, tools like credit might not make sense
link |
00:11:38.560
to people. Like it's a good example of something that does seem to pop up and be useful or just
link |
00:11:43.840
the power of markets and just in general capitalism. Yeah. And finance, I think the primary example
link |
00:11:49.280
I would give is leverage, right? So being allowed to borrow, to sort of use 10 times as much money
link |
00:11:56.880
as you've actually borrowed, right? So that's an example of something that before I had any
link |
00:12:00.880
experience in financial markets, I might have looked at and said, well, what is the purpose
link |
00:12:04.880
of that? That just seems very dangerous. And it is dangerous and it has proven dangerous.
link |
00:12:10.400
But, you know, if the fact of the matter is that, you know, sort of on some particular timescale,
link |
00:12:16.560
you are holding positions that are, you know, very unlikely to, you know,
link |
00:12:22.960
lose, you know, they're like your value at risk or variance is like one or five percent.
link |
00:12:29.120
Then it kind of makes sense that you would be allowed to use a little bit more than you have
link |
00:12:32.720
because you have, you know, some confidence that you're not going to lose it all in a single day.
link |
00:12:39.840
Now, of course, when that happens, we've seen what happens, you know, not too long ago. But,
link |
00:12:46.240
you know, but the idea that it serves no useful economic purpose under any circumstances is
link |
00:12:53.040
definitely not true. We'll return to the other side of the coast, Silicon Valley,
link |
00:12:58.640
and the problems there as we talk about, privacy as we talk about fairness.
link |
00:13:05.360
At the high level, and I'll ask some sort of basic questions with the hope to get at the
link |
00:13:10.560
fundamental nature of reality, but from a very high level, what is an ethical algorithm? So,
link |
00:13:18.320
I can say that an algorithm has a running time of using big old notation and log n.
link |
00:13:23.760
And I can say that a machine learning algorithm classified cat versus dog with 97% accuracy.
link |
00:13:31.440
Do you think there will one day be a way to measure sort of in the same compelling way as
link |
00:13:38.640
the big old notation of this algorithm is 97% ethical? First of all, let me riff for a second
link |
00:13:47.200
on your specific n log n examples. So, because early in the book when we're just kind of trying
link |
00:13:51.840
to describe algorithms, period, we say like, okay, you know, what's an example of an algorithm
link |
00:13:57.840
or an algorithmic problem? First of all, like it's sorting, right? You have a bunch of index
link |
00:14:01.840
cards with numbers on them and you want to sort them. And we describe, you know, an algorithm
link |
00:14:06.160
that sweeps all the way through finds that the smallest number puts it at the front then sweeps
link |
00:14:10.960
through again finds the second smallest number. So, we make the point that this is an algorithm,
link |
00:14:16.160
and it's also a bad algorithm in the sense that, you know, it's quadratic rather than n log n,
link |
00:14:22.080
which we know is kind of optimal for sorting. And we make the point that sort of like, you know,
link |
00:14:27.440
so even within the confines of a very precisely specified problem, there's, you know, there
link |
00:14:34.560
might be many, many different algorithms for the same problem with different properties,
link |
00:14:38.880
like some might be faster in terms of running time, some might use less memory, some might have,
link |
00:14:44.400
you know, better distributed implementations. And so, the point is, is that already we're used to,
link |
00:14:51.520
you know, in computer science thinking about tradeoffs between different types of quantities
link |
00:14:57.040
and resources, and there being, you know, better and worse algorithms. And our book is about that
link |
00:15:05.920
part of algorithmic ethics that we know how to kind of put on that same kind of quantitative footing
link |
00:15:14.240
right now. So, you know, just to say something that our book is not about, our book is not about
link |
00:15:20.160
kind of broad fuzzy notions of fairness. It's about very specific notions of fairness.
link |
00:15:27.360
There's more than one of them. There are tensions between them, right? But if you pick one of them,
link |
00:15:33.840
you can do something akin to saying that this algorithm is 97% ethical.
link |
00:15:40.560
You can say, for instance, the, you know, for this lending model, the false rejection rate
link |
00:15:47.360
on black people and white people is within 3%, right? So, we might call that a 97% ethical
link |
00:15:55.520
algorithm and a 100% ethical algorithm would mean that that difference is 0%.
link |
00:16:00.160
In that case, fairness is specified when two groups, however they're defined, are given to you.
link |
00:16:07.920
That's right. So, the, and then you can sort of mathematically start describing the algorithm.
link |
00:16:12.800
But nevertheless, the part where the two groups are given to you,
link |
00:16:20.960
I mean, unlike running time, you know, we don't, in computer science, talk about
link |
00:16:26.400
how fast an algorithm feels like when it runs. True. We measure it and ethical starts getting
link |
00:16:33.120
into feelings. So, for example, an algorithm runs, you know, if it runs in the background,
link |
00:16:39.600
it doesn't disturb the performance of my system. It'll feel nice. I'll be okay with it. But if
link |
00:16:44.320
it overloads the system, it'll feel unpleasant. So, in that same way, ethics, there's a feeling of
link |
00:16:50.160
how socially acceptable it is. How does it represent the moral standards of our society today? So,
link |
00:16:57.200
in that sense, and sorry to linger on that first of, high level philosophical question is,
link |
00:17:02.240
do you have a sense we'll be able to measure how ethical an algorithm is?
link |
00:17:07.440
First of all, I didn't, certainly didn't mean to give the impression that you can kind of measure,
link |
00:17:11.840
you know, memory, speed, tradeoffs, you know, and that there's a complete, you know, mapping
link |
00:17:17.600
from that onto kind of fairness, for instance, or ethics and accuracy, for example.
link |
00:17:24.400
In the type of fairness definitions that are largely the objects of study today and starting
link |
00:17:30.400
to be deployed, you as the user of the definitions, you need to make some hard decisions before you
link |
00:17:36.800
even get to the point of designing fair algorithms. One of them, for instance, is deciding who it is
link |
00:17:44.800
that you're worried about protecting, who you're worried about being harmed by, for instance,
link |
00:17:50.560
some notion of discrimination or unfairness. And then you need to also decide what constitutes harm.
link |
00:17:56.880
So, for instance, in a lending application, maybe you decide that, you know, falsely rejecting a
link |
00:18:03.680
credit worthy individual, you know, sort of a false negative is the real harm and that false
link |
00:18:10.320
positives, i.e., people that are not credit worthy or are not going to repay your loan,
link |
00:18:15.200
they get a loan, you might think of them as lucky. And so that's not a harm, although it's not clear
link |
00:18:20.720
that if you don't have the means to repay a loan that being given a loan is not also a harm.
link |
00:18:28.080
So, you know, the literature is sort of so far quite limited in that you sort of need to say,
link |
00:18:34.720
who do you want to protect and what would constitute harm to that group? And when you ask
link |
00:18:40.080
questions like, will algorithms feel ethical? One way in which they won't under the definitions
link |
00:18:45.680
that I'm describing is if, you know, if you are an individual who is falsely denied a loan,
link |
00:18:52.400
incorrectly denied a loan, all of these definitions basically say like, well,
link |
00:18:56.720
you know, your compensation is the knowledge that we are, we're also falsely denying loans to other
link |
00:19:02.880
people, you know, in other groups at the same rate that we're doing it to you. And, you know,
link |
00:19:08.640
and so there is actually this interesting even technical tension in the field right now between
link |
00:19:14.400
these sort of group notions of fairness and notions of fairness that might actually feel
link |
00:19:20.240
like real fairness to individuals, right? They might really feel like their particular interests
link |
00:19:25.920
are being protected or thought about by the algorithm rather than just, you know, the groups
link |
00:19:32.080
that they happen to be members of. Is there parallels to the big O notation of worst case
link |
00:19:38.000
analysis? So is it important to looking at the worst violation of fairness for an individual?
link |
00:19:46.400
Is it important to minimize that one individual? So like worst case analysis, is that something
link |
00:19:51.760
you think about or? I mean, I think we're not even at the point where we can sensibly think
link |
00:19:57.280
about that. So first of all, you know, we're talking here both about fairness applied at the
link |
00:20:03.120
group level, which is a relatively weak thing, but it's better than nothing. And also the more
link |
00:20:10.240
ambitious thing of trying to give some individual promises. But even that doesn't incorporate,
link |
00:20:17.120
I think something that you're hinting at here is what I might call subjective fairness, right?
link |
00:20:21.520
So a lot of the definitions, I mean, all of the definitions in the algorithmic fairness literature
link |
00:20:26.400
are what I would kind of call received wisdom definitions. It's sort of, you know, somebody
link |
00:20:31.120
like me sits around and things like, okay, you know, I think here's a technical definition
link |
00:20:35.360
of fairness that I think people should want or that they should, you know, think of as some
link |
00:20:40.240
notion of fairness, maybe not the only one, maybe not the best one, maybe not the last one.
link |
00:20:44.960
But we really actually don't know from a subjective standpoint, like what people really
link |
00:20:53.040
think is fair. There's, you know, we just started doing a little bit of work in our group
link |
00:20:59.200
at really actually doing kind of human subject experiments in which we, you know,
link |
00:21:06.080
ask people about, you know, we ask them questions about fairness, we survey them,
link |
00:21:12.320
we, you know, we show them pairs of individuals in let's say a criminal recidivism prediction
link |
00:21:17.440
setting. And we ask them, do you think these two individuals should be treated the same as a matter
link |
00:21:23.600
of fairness? And to my knowledge, there's not a large literature in which ordinary people
link |
00:21:30.720
are asked about, you know, they have sort of notions of their subjective fairness elicited
link |
00:21:36.800
from them. It's mainly, you know, kind of scholars who think about fairness, you know,
link |
00:21:42.960
kind of making up their own definitions. And I think, I think this needs to change actually
link |
00:21:46.880
for many social norms, not just for fairness, right? So there's a lot of discussion these days
link |
00:21:53.680
in the AI community about interpretable AI or understandable AI. And as far as I can tell,
link |
00:22:00.880
everybody agrees that deep learning or at least the outputs of deep learning are not very
link |
00:22:07.280
understandable. And people might agree that sparse linear models with integer coefficients are more
link |
00:22:14.400
understandable. But nobody's really asked people, you know, there's very little literature on, you
link |
00:22:19.200
know, sort of showing people models and asking them, do they understand what the model is doing?
link |
00:22:25.280
And I think that in all these topics, as these fields mature, we need to start doing more
link |
00:22:32.560
behavioral work. Yeah, which is so one of my deep passions is psychology. And I always thought
link |
00:22:39.360
computer scientists will be the best future psychologists in the sense that data is,
link |
00:22:48.160
especially in this modern world, the data is a really powerful way to understand and study
link |
00:22:52.160
human behavior. And you've explored that with your game theory side of work as well.
link |
00:22:56.720
Yeah, I'd like to think that what you say is true about computer scientists and psychology
link |
00:23:02.240
from my own limited wandering into human subject experiments. We have a great deal to learn.
link |
00:23:09.280
Not just computer science, but AI and machine learning, more specifically, I kind of think of
link |
00:23:13.280
as imperialist research communities in that, you know, kind of like physicists in an earlier
link |
00:23:19.280
generation, computer scientists kind of don't think of any scientific topic as off limits to
link |
00:23:25.280
them, they will like freely wander into areas that others have been thinking about for decades or
link |
00:23:31.040
longer. And, you know, we usually tend to embarrass ourselves in those efforts for some amount of
link |
00:23:38.800
time. Like, you know, I think reinforcement learning is a good example, right? So a lot of the early
link |
00:23:44.560
work in reinforcement learning, I have complete sympathy for the control theorists that looked
link |
00:23:50.480
at this and said like, okay, you are reinventing stuff that we've known since like the 40s, right?
link |
00:23:56.880
But, you know, in my view, eventually, this sort of, you know, computer scientists have made
link |
00:24:03.040
significant contributions to that field, even though we kind of embarrassed ourselves for the
link |
00:24:07.920
first decade. So I think if computer scientists are going to start engaging in kind of psychology,
link |
00:24:13.120
human subjects type of research, we should expect to be embarrassing ourselves for a good 10 years
link |
00:24:19.840
or so, and then hope that it turns out as well as, you know, some other areas that we've weighted
link |
00:24:25.600
into. So you kind of mentioned this, just to linger on the idea of an ethical algorithm,
link |
00:24:31.040
of idea of groups, sort of group thinking and individual thinking. And we're struggling that,
link |
00:24:35.920
that one of the amazing things about algorithms and your book and just this field of study is
link |
00:24:41.680
it gets us to ask, like, forcing machines, converting these ideas into algorithms is
link |
00:24:48.000
forcing us to ask questions of ourselves as a human civilization. So there's a lot of people now
link |
00:24:54.480
in public discourse doing sort of group thinking, thinking like there's particular
link |
00:25:00.640
sets of groups that we don't want to discriminate against and so on. And then there is
link |
00:25:04.320
individuals, sort of in the individual life stories, the struggles they went through, and so on.
link |
00:25:11.680
Now, like, in philosophy, it's easier to do group thinking because you don't,
link |
00:25:17.200
you know, it's very hard to think about individuals that there's so much variability.
link |
00:25:21.440
But with data, you can start to actually say, you know, what group thinking is too crude?
link |
00:25:28.080
You're actually doing more discrimination by thinking in terms of groups and individuals.
link |
00:25:32.560
Can you linger on that kind of idea of group versus individual and ethics? And is it good to
link |
00:25:40.080
continue thinking in terms of groups in algorithms? So let me start by answering a very good high
link |
00:25:47.600
level question with a slightly narrow technical response, which is these group definitions of
link |
00:25:54.080
fairness, like here's a few groups, like different racial groups, maybe gender groups, maybe age,
link |
00:25:59.440
what have you. And let's make sure that, you know, for none of these groups, do we, you know,
link |
00:26:06.720
have a false negative rate, which is much higher than any other one of these groups.
link |
00:26:10.400
Okay. So these are kind of classic group aggregate notions of fairness. And, you know, but at the
link |
00:26:16.000
end of the day, an individual you can think of as a combination of all their attributes, right?
link |
00:26:20.560
They're a member of a racial group, they have a gender, they have an age, you know, and many other,
link |
00:26:27.760
you know, demographic properties that are not biological, but that, you know, are, are still,
link |
00:26:33.280
you know, very strong determinants of outcome and personality and the like. So one, I think,
link |
00:26:39.680
useful spectrum is to sort of think about that array between the group and the specific individual
link |
00:26:46.400
and to realize that in some ways, asking for fairness at the individual level is to sort of
link |
00:26:52.400
ask for group fairness simultaneously for all possible combinations of groups. So in particular,
link |
00:27:00.000
so in particular, you know, if I build a predictive model that meets some definition of
link |
00:27:06.800
fairness by race, by gender, by age, by what have you, marginally to get it slightly technical,
link |
00:27:15.120
sort of independently, I shouldn't expect that model to not discriminate against disabled
link |
00:27:22.080
Hispanic women over age 55, making less than $50,000 a year annually, even though I might have
link |
00:27:28.000
protected each one of those attributes marginally. So the optimization, actually, that's a fascinating
link |
00:27:34.640
way to put it. So you're just optimizing the one way to achieve the optimizing fairness for
link |
00:27:41.600
individuals, just add more and more definitions of groups that each individual belongs to.
link |
00:27:46.000
So, you know, at the end of the day, we could think of all of ourselves as groups of size one,
link |
00:27:50.800
because eventually there's some attribute that separates you from me and everybody,
link |
00:27:54.640
from everybody else in the world. Okay. And so it is possible to put, you know,
link |
00:28:00.960
these incredibly coarse ways of thinking about fairness and these very, very individualistic
link |
00:28:06.080
specific ways on a common scale. And, you know, one of the things we've worked on from a research
link |
00:28:12.560
perspective is, you know, so we sort of know how to, you know, we, in relative terms, we know
link |
00:28:18.000
how to provide fairness guarantees at the course ascend of the scale. We don't know how to provide
link |
00:28:24.240
kind of sensible, tractable, realistic fairness guarantees at the individual level. But maybe
link |
00:28:29.760
we could start creeping towards that by dealing with more, you know, refined subgroups. I mean,
link |
00:28:35.120
we gave a name to this phenomenon where, you know, you protect, you enforce some definition of
link |
00:28:42.160
fairness for a bunch of marginal attributes or features, but then you find yourself discriminating
link |
00:28:48.160
against a combination of them. We call that fairness gerrymandering, because like political
link |
00:28:53.520
gerrymandering, you know, you're giving some guarantee at the aggregate level. But when you
link |
00:28:59.760
kind of look in a more granular way at what's going on, you realize that you're achieving
link |
00:29:04.000
that aggregate guarantee by sort of favoring some groups and discriminating against other ones.
link |
00:29:09.200
And so there are, you know, it's early days, but there are algorithmic approaches that let you
link |
00:29:15.520
start creeping towards that, you know, individual end of the spectrum.
link |
00:29:21.360
Does there need to be human input in the form of weighing the value of the importance of each kind
link |
00:29:30.720
of group? So, for example, is it, is it like, so gender, say, crudely speaking, male and female,
link |
00:29:41.040
and then different races, are we as humans supposed to put value on saying gender is 0.6
link |
00:29:50.080
and race is 0.4 in terms of in the big optimization of achieving fairness? Is that kind of what humans
link |
00:29:59.600
are supposed to do here? I mean, of course, you know, I don't need to tell you that, of course,
link |
00:30:04.400
technically, one could incorporate such weights if you wanted to into a definition of fairness.
link |
00:30:11.280
You know, fairness is an interesting topic in that having worked in the book being about
link |
00:30:19.600
both fairness, privacy, and many other social norms. Fairness, of course, is a much,
link |
00:30:24.560
much more loaded topic. So, privacy, I mean, people want privacy. People don't like violations of
link |
00:30:31.120
privacy. Violations of privacy cause damage, angst, and bad publicity for the companies that
link |
00:30:38.560
are victims of them. But sort of everybody agrees, more data privacy would be better
link |
00:30:45.200
than less data privacy. And you don't have these, somehow the discussions of fairness don't become
link |
00:30:51.760
politicized along other dimensions like race and about gender and, you know, whether we,
link |
00:31:00.320
and, you know, you quickly find yourself kind of revisiting topics that have been
link |
00:31:08.320
kind of unresolved forever, like affirmative action, right? So, you know, like, why are you
link |
00:31:14.160
protecting, you know, some people will say, why are you protecting this particular racial group?
link |
00:31:17.840
And, and others will say, well, we need to do that as a matter of, of retribution. Other people
link |
00:31:26.560
will say it's a matter of economic opportunity. And I don't know which of, you know, whether any
link |
00:31:33.680
of these are the right answers, but you sort of, fairness is sort of special in that as soon as
link |
00:31:37.760
you start talking about it, you inevitably have to participate in debates about fair to whom,
link |
00:31:44.880
at what expense, to who else. I mean, even in criminal justice, right? You know, where people
link |
00:31:52.240
talk about fairness in criminal sentencing or, you know, predicting failures to appear or making
link |
00:32:01.840
parole decisions or the like, they will, you know, they'll point out that, well, these definitions
link |
00:32:08.160
of fairness are all about fairness for the criminals. And what about fairness for the victims,
link |
00:32:15.440
right? So when I basically say something like, well, the, the false incarceration rate for
link |
00:32:22.160
black people and white people needs to be roughly the same. You know, there's no mention of potential
link |
00:32:28.240
victims of criminals in such a fairness definition. And that's the realm of public discord. I should
link |
00:32:35.280
actually recommend, I just listened to, to people listening, Intelligent Squares Debates, U.S.
link |
00:32:42.400
Edition just had a debate. They have this structure where you have an old Oxford style or whatever
link |
00:32:49.040
they're called, debates, those two versus two, and they talked about affirmative action. And it was,
link |
00:32:54.720
it was incredibly interesting that it's still, there's really good points on every side of this
link |
00:33:01.520
issue, which is fascinating to listen to. Yeah, yeah, I agree. And so it's, it's interesting to be
link |
00:33:07.760
a researcher trying to do, for the most part, technical algorithmic work. But Aaron and I
link |
00:33:14.880
both quickly learned, you cannot do that and then go out and talk about it and expect people to take
link |
00:33:19.520
it seriously. If you're unwilling to engage in these broader debates that are, are entirely
link |
00:33:25.760
extra algorithmic, right? They're, they're, they're not about, you know, algorithms and making
link |
00:33:30.240
algorithms better. They're sort of, you know, as you said, sort of like, what should society be
link |
00:33:34.560
protecting in the first place? When you discuss a fairness, an algorithm that, that achieves fairness,
link |
00:33:41.120
whether in the constraints and the objective function, there's an immediate kind of analysis
link |
00:33:46.240
you can perform, which is saying, if you care about fairness in gender, this is the amount
link |
00:33:54.080
that you have to pay for it in terms of the performance of the system. Like, do you, is there
link |
00:34:00.080
a role for statements like that in a table, in a paper, or do you want to really not touch that?
link |
00:34:06.320
Like, no, we, we want to touch that and we do touch it. So I mean, just, just again, to make sure
link |
00:34:12.320
I'm not promising your, your viewers more than we know how to provide. But if you pick a definition
link |
00:34:18.640
of fairness, like I'm worried about gender discrimination and you pick a notion of harm,
link |
00:34:23.280
like false rejection for a loan, for example, and you give me a model, I can definitely,
link |
00:34:29.120
first of all, go audit that model. It's easy for me to go, you know, from data to kind of say,
link |
00:34:34.880
like, okay, your false rejection rate on women is this much higher than it is on men. Okay.
link |
00:34:41.520
But, you know, once you also put the fairness into your objective function, I mean, I think the
link |
00:34:46.800
table that you're talking about is, you know, what, what we would call the Pareto curve,
link |
00:34:50.800
right? You can literally trace out and we give examples of such plots on real data sets in the
link |
00:34:57.760
book. You have two axes on the x axis is your error on the y axis is unfairness by whatever,
link |
00:35:05.680
you know, if it's like the disparity between false rejection rates between two groups.
link |
00:35:11.680
And, you know, your algorithm now has a knob that basically says, how strongly do I want
link |
00:35:16.960
to enforce fairness? And the less unfair, you know, we, you know, if the two axes are
link |
00:35:22.880
error and unfairness, we'd like to be at zero zero, we'd like zero error and zero fair unfairness
link |
00:35:29.120
simultaneously. Anybody who works in machine learning knows that you're generally not going
link |
00:35:34.400
to get to zero error period without any fairness constraint whatsoever. So that that that's not
link |
00:35:39.840
going to happen. But in general, you know, you'll get this, you'll get some kind of convex curve
link |
00:35:46.240
that specifies the numerical tradeoff you face, you know, if I want to go from 17%
link |
00:35:53.360
error down to 16% error, what will be the increase in unfairness that I would experience as a result
link |
00:36:00.560
of that. And, and so this curve kind of specifies the, you know, kind of
link |
00:36:07.120
undominated models, models that are off that curve are, you know, can be strictly improved in one or
link |
00:36:13.520
both dimensions, you can, you know, either make the error better or the unfairness better or both.
link |
00:36:18.560
And I think our view is that not only are these objects, these Pareto curves, you know,
link |
00:36:25.760
they're efficient frontiers, as you might call them. Not only are they valuable scientific
link |
00:36:33.920
objects, I actually think that they in the near term might need to be the interface between
link |
00:36:41.360
researchers working in the field and, and stakeholders in given problems. So, you know,
link |
00:36:46.960
you could really imagine telling a criminal jurisdiction, look, if you're concerned about
link |
00:36:55.600
racial fairness, but you're also concerned about accuracy, you want to, you know, you want to
link |
00:37:01.840
release on parole people that are not going to recommit a violent crime and you don't want to
link |
00:37:06.720
release the ones who are. So, you know, that's accuracy. But if you also care about those,
link |
00:37:12.320
you know, the mistakes you make not being disproportionately on one racial group or another,
link |
00:37:17.120
you can, you can show this curve. I'm hoping that in the near future, it'll be possible to
link |
00:37:21.920
explain these curves to non technical people that have that are the ones that have to make the
link |
00:37:27.680
decision. Where do we want to be on this curve? Like what are the relative merits or value of
link |
00:37:34.880
having lower air versus lower unfairness? You know, that's not something computer scientists
link |
00:37:41.280
should be deciding for society, right? That, you know, the people in the field, so to speak,
link |
00:37:46.720
the policy makers, the regulators, that's who should be making these decisions.
link |
00:37:51.360
But I think and hope that they can be made to understand that these tradeoffs generally exist
link |
00:37:57.840
and that you need to pick a point and like, and ignoring the tradeoff, you know,
link |
00:38:03.040
you're implicitly picking a point anyway, right? You just don't know it and you're not admitting it.
link |
00:38:09.120
Just to link down the point of tradeoffs, I think that's a really important thing to sort of
link |
00:38:14.560
think about. So you think when we start to optimize for fairness, there's almost always
link |
00:38:21.440
in most system going to be tradeoffs. So can you like, what's the tradeoff between just to
link |
00:38:28.240
clarify? There have been some sort of technical terms thrown around, but sort of
link |
00:38:36.560
a perfectly fair world. Why is that? Why will somebody be upset about that?
link |
00:38:43.520
The specific tradeoff I talked about just in order to make things very concrete was between
link |
00:38:48.640
numerical error and some numerical measure of unfairness.
link |
00:38:52.160
What is numerical error in the case of? Just like say predictive error, like, you know,
link |
00:38:57.840
the probability or frequency with which you release somebody on parole who then goes on to
link |
00:39:04.240
recommit of violent crime or keep incarcerated somebody who would not have recommitted of
link |
00:39:09.920
violent crime. So in the case of awarding somebody parole or giving somebody parole or letting them
link |
00:39:16.880
out on parole, you don't want them to recommit of crime. So it's your system failed in prediction
link |
00:39:23.760
if they happen to do a crime. Okay. So that's the performance. That's one axis.
link |
00:39:29.280
Right. And what's the fairness axis? So then the fairness axis might be the difference between
link |
00:39:34.480
racial groups in the kind of false, false positive predictions, namely people that I
link |
00:39:41.920
kept incarcerated predicting that they would recommit of violent crime when in fact they
link |
00:39:48.640
wouldn't have. Right. And the unfairness of that, just to linger it, and allow me to
link |
00:39:56.800
ineliquently to try to sort of describe why that's unfair, why unfairness is there.
link |
00:40:03.040
The, the unfairness you want to get rid of is that in the judge's mind, the bias of having
link |
00:40:10.560
being brought up to society, the slight racial bias, the racism that exists in the society,
link |
00:40:15.520
you want to remove that from the system. Another way that's been debated is sort of equality
link |
00:40:24.080
of opportunity versus equality of outcome. And there's a weird dance there that's
link |
00:40:30.720
really difficult to get right. And we don't, the affirmative action is exploring that space.
link |
00:40:37.920
Right. And then we, this also quickly, you know, bleeds into questions like, well,
link |
00:40:44.560
maybe if one group really does recommit crimes at a higher rate, the reason for that is that at
link |
00:40:51.440
some earlier point in the pipeline or earlier in their lives, they didn't receive the same
link |
00:40:56.080
resources that the other group did. Right. And that, and so, you know, there's always in, in
link |
00:41:02.160
kind of fairness discussions, the possibility that the, the real injustice came earlier,
link |
00:41:07.680
right, earlier in this individual's life, earlier in this group's history, etc., etc.
link |
00:41:13.440
And so a lot of the fairness discussion is almost, the goal is for it to be a corrective
link |
00:41:18.480
mechanism to account for the injustice earlier in life.
link |
00:41:22.560
By some definitions of fairness or some theories of fairness, yeah. Others would say, like, look,
link |
00:41:27.840
it's, it's, you know, it's not to correct that injustice, it's just to kind of level the playing
link |
00:41:32.880
field right now and not incarcerate, falsely incarcerate more people of one group than another
link |
00:41:38.320
group. But I mean, do you think just, it might be helpful just to demystify a little bit about
link |
00:41:43.440
the many ways in which bias or unfairness can come into
link |
00:41:49.440
algorithms, especially in the machine learning era, right. And, you know, I think many of your
link |
00:41:53.680
viewers have probably heard these examples before. But, you know, let's say I'm building a face
link |
00:41:58.320
recognition system, right. And so, you know, kind of gathering lots of images of faces and,
link |
00:42:04.400
you know, trying to train the system to, you know, recognize new faces of those individuals
link |
00:42:10.400
from training on, you know, a training set of those faces of individuals. And, you know,
link |
00:42:15.840
it shouldn't surprise anybody, or certainly not anybody in the field of machine learning,
link |
00:42:20.560
if my training dataset was primarily white males, and I'm training that the model to
link |
00:42:26.640
maximize the overall accuracy on my training dataset, that, you know, the model can reduce
link |
00:42:35.040
its error most by getting things right on the white data set. And, you know,
link |
00:42:40.720
the model can reduce its error most by getting things right on the white males that constitute
link |
00:42:46.880
the majority of the data set, even if that means that on other groups, they will be less accurate,
link |
00:42:52.560
okay. Now, there's a bunch of ways you could think about addressing this. One is to deliberately
link |
00:42:58.800
put into the objective of the algorithm not to, not to optimize the error at the expense of
link |
00:43:06.560
this discrimination. And then you're kind of back in the land of these kind of two dimensional
link |
00:43:10.000
numerical tradeoffs. A valid counterargument is to say like, well, no, you don't have to,
link |
00:43:16.560
there's no, you know, the notion of the tension between error and accuracy here is a false one.
link |
00:43:22.560
You could instead just go out and get much more data on these other groups that are in the minority
link |
00:43:29.040
and, you know, equalize your dataset, or you could train a separate model on those subgroups and,
link |
00:43:35.760
you know, have multiple models. The point I think we would, you know, we tried to make in the book
link |
00:43:41.760
is that those things have cost too, right? Going out and gathering more data on groups that are
link |
00:43:49.280
relatively rare compared to your plurality or more majority group that, you know, it may not
link |
00:43:54.240
cost you in the accuracy of the model, but it's going to cost, you know, it's going to cost the
link |
00:43:58.480
company developing this model more money to develop that. And it also costs more money to build
link |
00:44:03.920
separate predictive models and to implement and deploy them. So even if you can find a way to
link |
00:44:10.000
avoid the tension between error and accuracy in training a model, you might push the cost somewhere
link |
00:44:17.200
else like money, like development time, research time and the like.
link |
00:44:22.640
There are fundamentally difficult philosophical questions in fairness. And we live in a very
link |
00:44:30.160
divisive political climate, outraged culture. There is alt right folks on 4chan trolls. There is
link |
00:44:40.240
social justice warriors on Twitter. There is very divisive, outraged folks on all sides of
link |
00:44:47.840
every kind of system. How do you, how do we as engineers build ethical algorithms in such
link |
00:44:56.080
divisive culture? Do you think they could be disjoint? The human has to inject your values
link |
00:45:02.080
and then you can optimize over those values. But in our times when, when you start actually
link |
00:45:07.680
applying these systems, things get a little bit challenging for the public discourse. How do you
link |
00:45:13.280
think we can proceed? Yeah, I mean, for the most part in the book, you know, a point that we
link |
00:45:18.640
try to take some pains to make is that we don't view ourselves or people like us as being in
link |
00:45:28.160
the position of deciding for society what the right social norms are, what the right definitions
link |
00:45:33.520
of fairness are. Our main point is to just show that if society or the relevant stakeholders
link |
00:45:40.880
in a particular domain can come to agreement on those sorts of things, there's a way of encoding
link |
00:45:46.720
that into algorithms in many cases, not in all cases. One other misconception that hopefully we
link |
00:45:52.720
definitely dispel is sometimes people read the title of the book and I think not unnaturally
link |
00:45:58.720
fear that what we're suggesting is that the algorithms themselves should decide what those
link |
00:46:03.280
social norms are and develop their own notions of fairness and privacy or ethics. And we're
link |
00:46:08.320
definitely not suggesting that. The title of the book is ethical algorithm, by the way, and I didn't
link |
00:46:12.640
think of that interpretation of the title. That's interesting. Yeah, yeah. I mean, especially these
link |
00:46:17.040
days where people are concerned about the robots becoming our overlords, the idea that the robots
link |
00:46:23.200
would also develop their own social norms is just one step away from that. But I do think,
link |
00:46:30.400
obviously, despite disclaimer that people like us shouldn't be making those decisions for society,
link |
00:46:36.400
we are living in a world where, in many ways, computer scientists have made some decisions
link |
00:46:42.720
that have fundamentally changed the nature of our society and democracy and civil discourse
link |
00:46:49.200
and deliberation in ways that I think most people generally feel are bad these days.
link |
00:46:55.520
But if we look at people at the heads of companies and so on, they had to make those
link |
00:47:01.600
decisions. There has to be decisions. So there's two options. Either you kind of put your head in
link |
00:47:08.560
the sand and don't think about these things and just let the algorithm do what it does,
link |
00:47:13.760
or you make decisions about what you value, injecting moral values into the algorithm.
link |
00:47:19.840
Look, I never mean to be an apologist for the tech industry, but I think it's a little bit too
link |
00:47:27.920
far to say that explicit decisions were made about these things. So let's, for instance,
link |
00:47:31.920
take social media platforms. So like many inventions in technology and computer science,
link |
00:47:38.240
a lot of these platforms that we now use regularly kind of started as curiosities.
link |
00:47:44.960
I remember when things like Facebook came out and its predecessors like Friendster,
link |
00:47:49.040
which nobody even remembers now, people really wonder why would anybody want to spend time
link |
00:47:55.840
doing that? I mean, even the web when it first came out, when it wasn't populated with much
link |
00:48:00.800
content and it was largely kind of hobbyists building their own kind of ramshackle websites,
link |
00:48:06.720
a lot of people looked at this as like, what is the purpose of this thing? Why is this
link |
00:48:10.320
interesting? Who would want to do this? And so even things like Facebook and Twitter,
link |
00:48:14.960
yes, technical decisions were made by engineers, by scientists, by executives,
link |
00:48:19.840
in the design of those platforms. But I don't think 10 years ago, anyone anticipated
link |
00:48:29.520
that those platforms, for instance, might kind of acquire undue influence on political discourse
link |
00:48:39.120
or on the outcomes of elections. And I think the scrutiny that these companies are getting now
link |
00:48:46.080
is entirely appropriate. But I think it's a little too harsh to kind of look at history
link |
00:48:51.840
and sort of say like, oh, you should have been able to anticipate that this would happen with
link |
00:48:55.200
your platform. And in the sort of gaming chapter of the book, one of the points we're making is that,
link |
00:49:00.400
you know, these platforms, right, they don't operate in isolation. So unlike the other topics
link |
00:49:06.480
we're discussing like fairness and privacy, like those are really cases where algorithms can operate
link |
00:49:11.600
on your data and make decisions about you. And you're not even aware of it. Okay. Things like
link |
00:49:16.400
Facebook and Twitter, these are, you know, these are, these are systems, right? These are social
link |
00:49:20.880
systems. And their evolution, even their technical evolution, because machine learning is involved,
link |
00:49:27.200
is driven in no small part by the behavior of the users themselves and how the users decide to adopt
link |
00:49:33.440
them and how to use them. And so, you know, you know, I'm kind of like, who really knew that,
link |
00:49:42.240
that, you know, until until we saw it happen, who knew that these things might be able to influence
link |
00:49:46.480
the outcome of elections? Who knew that, you know, they might polarize political discourse because
link |
00:49:53.840
of the ability to, you know, decide who you interact with on the platform and also with the
link |
00:49:59.920
platform naturally using machine learning to optimize for your own interests that they would
link |
00:50:04.720
further isolate us from each other and, you know, like feed us all basically just the stuff that
link |
00:50:10.160
we already agreed with. And so I think, you know, we've come to that outcome, I think, largely,
link |
00:50:15.760
but I think it's something that we all learned together, including the companies as these things
link |
00:50:22.720
happen. Now, you asked like, well, are there algorithmic remedies to these kinds of things? And
link |
00:50:31.120
again, these are big problems that are not going to be solved with, you know,
link |
00:50:34.960
somebody going in and changing a few lines of code somewhere in a social media platform.
link |
00:50:39.760
But I do think in many ways, there are definitely ways of making things better. I mean,
link |
00:50:45.040
like an obvious recommendation that we make at some point in the book is like, look, you know,
link |
00:50:49.360
to the extent that we think that machine learning applied for personalization purposes in things
link |
00:50:55.760
like news feed, you know, or other platforms has led to polarization and intolerance of opposing
link |
00:51:04.640
viewpoints. As you know, right, these, these algorithms have models, right? And they kind of
link |
00:51:10.400
place people in some kind of metric space and, and they place content in that space. And they
link |
00:51:15.600
sort of know the extent to which I have an affinity for a particular type of content. And by the
link |
00:51:21.280
same token, they also probably have that same model probably gives you a good idea of the stuff
link |
00:51:27.120
I'm likely to violently disagree with or be offended by. Okay. So, you know, in this case,
link |
00:51:33.520
there really is some knob you could tune that says like, instead of showing people only what they
link |
00:51:38.960
like and what they want, let's show them some stuff that we think that they don't like, or that's a
link |
00:51:43.760
little bit further away. And you could even imagine users being able to control this, you know, just
link |
00:51:49.040
like a, everybody gets a slider. And that slider says like, you know, how much stuff do you want
link |
00:51:55.920
to see that's kind of, you know, you might disagree with, or is at least further from your
link |
00:52:00.960
interest. Like, it's almost like an exploration button. So just get your intuition. Do you think
link |
00:52:08.000
do you think engagement? So like, you're staying on the platform, you're staying engaged. Do you
link |
00:52:14.000
think fairness, ideas of fairness won't emerge? Like, how bad is it to just optimize for engagement?
link |
00:52:22.400
Do you think we'll run into big trouble if we're just optimizing for how much you love the platform?
link |
00:52:28.640
Well, I mean, optimizing for engagement kind of got us where we are.
link |
00:52:32.720
So, do you, one, have faith that it's possible to do better? And two, if it is, how do we do better?
link |
00:52:42.240
I mean, it's definitely possible to do different, right? And again, you know, it's not as if I think
link |
00:52:47.440
that doing something different than optimizing for engagement won't cost these companies in real
link |
00:52:53.200
ways, including revenue and profitability potentially. In the short term, at least.
link |
00:52:58.640
Yeah, in the short term, right. And again, you know, if I worked at these companies, I'm sure that
link |
00:53:08.160
it would have seemed like the most natural thing in the world also to want to optimize
link |
00:53:11.680
engagement, right? And that's good for users in some sense. You want them to be, you know,
link |
00:53:16.160
vested in the platform and enjoying it and finding it useful, interesting and or productive.
link |
00:53:21.360
But, you know, my point is, is that the idea that there is, that it's sort of out of their hands,
link |
00:53:27.440
as you said, or that there's nothing to do about it, never say never, but that strikes
link |
00:53:31.840
me as implausible as a machine learning person, right? I mean, these companies are driven by
link |
00:53:36.240
machine learning and this optimization of engagement is essentially driven by machine
link |
00:53:41.280
learning, right? It's driven by not just machine learning, but, you know, very, very large scale
link |
00:53:46.720
A, B experimentation where you kind of tweak some element of the user interface or tweak some
link |
00:53:52.960
component of an algorithm or tweak some component or feature of your click through
link |
00:53:58.960
prediction model. And my point is, is that anytime you know how to optimize for something,
link |
00:54:05.200
you know, by def, almost by definition, that solution tells you how not to optimize for it
link |
00:54:10.320
or to do something different. Engagement can be measured. So, sort of optimizing
link |
00:54:17.600
for, sort of minimizing divisiveness or maximizing intellectual growth over the lifetime of a human
link |
00:54:26.880
being are very difficult to measure. That's right. So, I'm not, I'm not claiming that doing
link |
00:54:33.280
something different will immediately make it apparent that this is a good thing for society.
link |
00:54:40.400
And in particular, I mean, I think one way of thinking about where we are on some of these
link |
00:54:44.640
social media platforms is that, you know, it kind of feels a bit like we're in a bad equilibrium,
link |
00:54:50.000
right? That these systems are helping us all kind of optimize something myopically and selfishly
link |
00:54:55.920
for ourselves. And of course, from an individual standpoint at any given moment, like, why would
link |
00:55:02.240
I want to see things in my newsfeed that I found irrelevant, offensive, or, you know, or the like,
link |
00:55:07.840
okay? But, you know, maybe by all of us, you know, having these platforms myopically optimize in our
link |
00:55:15.600
interests, we have reached a collective outcome as a society that we're unhappy with in different
link |
00:55:21.280
ways, let's say, with respect to things like, you know, political discourse and tolerance
link |
00:55:26.160
of opposing viewpoints. And if Mark Zuckerberg gave you a call and said, I'm thinking of taking
link |
00:55:33.920
a sabbatical, could you run Facebook for me for six months? What would you how?
link |
00:55:38.240
I think no thanks would be the first response. But there are many aspects of being the head of
link |
00:55:45.840
the entire company that are kind of entirely exogenous to many of the things that we're discussing
link |
00:55:51.680
here. And so I don't really think I would need to be CEO of Facebook to kind of implement the,
link |
00:55:58.480
you know, more limited set of solutions that I might imagine. But I think one, one concrete
link |
00:56:04.880
thing they could do is they could experiment with letting people who chose to, to see more stuff in
link |
00:56:12.560
their newsfeed that is not entirely kind of chosen to optimize for their particular interests,
link |
00:56:20.240
beliefs, etc. So the kind of thing is like I speak to YouTube, but I think Facebook probably
link |
00:56:27.200
does something similar is they're quite effective at automatically finding what sorts of groups you
link |
00:56:34.960
belong to, not based on race or gender or so on, but based on the kind of stuff you enjoy watching
link |
00:56:41.360
in the case of YouTube. So it's a difficult thing for Facebook or YouTube to then say,
link |
00:56:49.760
well, you know what, we're going to show you something from a very different cluster.
link |
00:56:53.120
Even though we believe algorithmically, you're unlikely to enjoy that thing.
link |
00:56:59.360
Sort of, that's a weird jump to make. There has to be a human, like at the very top of
link |
00:57:04.960
that system that says, well, that will be long term healthy for you. That's more than an algorithmic
link |
00:57:11.040
decision. Or that same person could say that'll be long term healthy for the platform or for
link |
00:57:17.520
the platform's influence on society outside of the platform. It's easy for me to sit here and say
link |
00:57:24.800
these things, but conceptually, I do not think that these are totally or they shouldn't be completely
link |
00:57:32.960
alien ideas. You could try things like this and we wouldn't have to invent entirely new science
link |
00:57:42.880
to do it because if we're all already embedded in some metric space and there's a notion of
link |
00:57:47.760
distance between you and me and every piece of content, then the same model that dictates how
link |
00:57:59.200
to make me really happy also tells how to make me as unhappy as possible as well.
link |
00:58:05.760
The focus in your book and algorithmic fairness research today in general is on machine learning,
link |
00:58:11.040
like we said, is data. Just even the entire AI field right now is captivated with machine
link |
00:58:17.280
learning, with deep learning. Do you think ideas in symbolic AI or totally other kinds of approaches
link |
00:58:24.320
are interesting, useful in the space, have some promising ideas in terms of fairness?
link |
00:58:31.200
I haven't thought about that question specifically in the context of fairness. I definitely would
link |
00:58:36.800
agree with that statement in the large. I am one of many machine learning researchers who
link |
00:58:45.360
do believe that the great successes that have been shown in machine learning recently
link |
00:58:50.160
are great successes, but they're on a pretty narrow set of tasks. I don't think we're
link |
00:58:56.800
notably closer to general artificial intelligence now than we were when I started my career.
link |
00:59:03.200
I mean, there's been progress. I do think that we are kind of as a community maybe looking at
link |
00:59:08.960
that where the light is, but the light is shining pretty bright there right now and we're finding
link |
00:59:12.880
a lot of stuff. I don't want to argue with the progress that's been made in areas like deep
link |
00:59:18.000
learning, for example. This touches another related thing that you mentioned and that people
link |
00:59:23.840
might misinterpret from the title of your book, Ethical Algorithm. Is it possible for the algorithm
link |
00:59:28.880
to automate some of those decisions, higher level decisions of what should be fair?
link |
00:59:36.720
What should be fair?
link |
00:59:38.480
The more you know about a field, the more aware you are of its limitations.
link |
00:59:43.600
I'm pretty leery of trying. There's so much we already don't know in fairness,
link |
00:59:51.600
even when we're the ones picking the fairness definitions and comparing alternatives and
link |
00:59:56.960
thinking about the tensions between different definitions, that the idea of kind of letting
link |
01:00:01.600
the algorithm start exploring as well. I definitely think this is a much narrower statement. I
link |
01:00:08.560
definitely think that kind of algorithmic auditing for different types of unfairness,
link |
01:00:12.560
right? So like in this gerrymandering example where I might want to prevent not just discrimination
link |
01:00:18.560
against very broad categories, but against combinations of broad categories, you quickly
link |
01:00:24.480
get to a point where there's a lot of categories. There's a lot of combinations of end features.
link |
01:00:31.920
You can use algorithmic techniques to sort of try to find the subgroups on which you're
link |
01:00:36.400
discriminating the most and try to fix that. That's actually kind of the form of one of the
link |
01:00:40.640
algorithms we developed for this fairness gerrymandering problem. But partly because of our
link |
01:00:48.080
techno law, our sort of our scientific ignorance on these topics right now. And also partly just
link |
01:00:54.240
because these topics are so loaded emotionally for people that I just don't see the value.
link |
01:01:00.640
I mean, again, never say never, but I just don't think we're at a moment where it's a great time
link |
01:01:04.560
for computer scientists to be rolling out the idea like, hey, not only have we kind of figured
link |
01:01:10.400
fairness out, but we think the algorithms should start deciding what's fair or giving input on
link |
01:01:16.240
that decision. I just don't. It's like the cost benefit analysis to the field of kind of going
link |
01:01:21.840
there right now just doesn't seem worth it to me. That said, I should say that I think computer
link |
01:01:26.880
scientists should be more philosophically like should enrich their thinking about these kinds
link |
01:01:31.440
of things. I think it's been too often used as an excuse for roboticists working autonomous vehicles,
link |
01:01:37.840
for example, to not think about the human factor or psychology or safety in the same way like computer
link |
01:01:44.480
sizes and algorithms that have been sort of using as an excuse. And I think it's time for basically
link |
01:01:49.360
everybody to become computer scientists. I was about to agree with everything you said except
link |
01:01:53.440
that last point. I think that the other way of looking at it is that I think computer scientists,
link |
01:01:59.200
you know, and many of us are, but we need to wade out into the world more, right? I mean,
link |
01:02:07.040
just the influence that computer science and therefore computer scientists have had on society
link |
01:02:13.280
at large, just like has exponentially magnified in the last 10 or 20 years or so. And, you know,
link |
01:02:22.640
you know, before when we were just tinkering around amongst ourselves and it didn't matter that much,
link |
01:02:27.600
there was no need for sort of computer scientists to be citizens of the world more broadly.
link |
01:02:33.120
And I think those days need to be over very, very fast. And I'm not saying everybody needs to do it,
link |
01:02:38.800
but to me, like the right way of doing it is to not to sort of think that everybody else is going
link |
01:02:42.560
to become a computer scientist. But, you know, I think, you know, people are becoming more
link |
01:02:47.280
sophisticated about computer science, even lay people. You know, I think one of the reasons we
link |
01:02:53.120
decided to write this book is we thought 10 years ago, I wouldn't have tried this just because I
link |
01:02:58.400
just didn't think that sort of people's awareness of algorithms and machine learning,
link |
01:03:04.000
you know, the general population would have been high. And I mean, you would have had to first,
link |
01:03:08.160
you know, write one of the many books kind of just explicating that topic to a lay audience first.
link |
01:03:14.480
Now I think we're at the point where like lots of people without any technical training
link |
01:03:18.800
at all know enough about algorithms and machine learning that you can start
link |
01:03:22.160
getting to these nuances of things like ethical algorithms. I think we agree that there needs to
link |
01:03:27.680
be much more mixing. But I think I think a lot of the onus of that mixing needs to be on the
link |
01:03:34.160
computer science community. Yeah. So just to linger on the disagreement, because I do disagree with
link |
01:03:39.760
you on the point that I think if you're a biologist, if you're a chemist, if you're an MBA business
link |
01:03:49.040
person, all of those things you can like if you learned a program and not only program, if you
link |
01:03:56.720
learned to do machine learning, if you learned to do data science, you immediately become much
link |
01:04:01.920
more powerful in the kinds of things you can do. And therefore, literature, like library sciences,
link |
01:04:07.680
like, so you were speaking, I think, I think it holds true what you're saying for the next
link |
01:04:13.760
two years. But long term, if you're interested to me, if you're interested in philosophy,
link |
01:04:21.040
you should learn a program. Because then you can scrape data, you can study what people are
link |
01:04:27.040
thinking about on Twitter, and then start making philosophical conclusions about the meaning
link |
01:04:32.400
of life. I just feel like the access to data, the digitization of whatever problem you're trying to
link |
01:04:39.920
solve, it fundamentally changes what it means to be a computer scientist. To me, a computer
link |
01:04:44.880
scientist in 20, 30 years will go back to being a Donald Knuth style theoretical computer science,
link |
01:04:51.760
and everybody would be doing basically, they're exploring the kinds of ideas that you're exploring
link |
01:04:56.960
in your book. It won't be a computer science. Yeah, I mean, I don't think I disagree enough,
link |
01:05:01.360
but I think that that trend of more and more people and more and more disciplines,
link |
01:05:08.320
adopting ideas from computer science, learning how to code, I think that that trend seems
link |
01:05:13.600
firmly underway. I mean, you know, like an interesting digressive question along these
link |
01:05:19.040
lines is, maybe in 50 years, there won't be computer science departments anymore,
link |
01:05:24.800
because the field will just sort of be ambient in all of the different disciplines. And people
link |
01:05:31.760
look back and having a computer science department will look like having an electricity department
link |
01:05:36.880
or something. It's like, everybody uses this, it's just out there. I mean, I do think there will
link |
01:05:41.920
always be that kind of Knuth style core to it. But it's not an implausible path that we kind of
link |
01:05:47.280
get to the point where the academic discipline of computer science becomes somewhat marginalized
link |
01:05:53.040
because of its very success in infiltrating all of science and society and the humanities, etc.
link |
01:06:00.400
What is differential privacy or more broadly algorithmic privacy?
link |
01:06:07.120
Algorithmic privacy more broadly is just the study or the notion of privacy
link |
01:06:12.880
definitions or norms being encoded inside of algorithms. And so, you know, I think we count
link |
01:06:21.600
among this body of work, just, you know, the literature and practice of things like data
link |
01:06:28.560
anonymization, which we kind of at the beginning of our discussion of privacy, say like, okay,
link |
01:06:35.200
this is this is sort of a notion of algorithmic privacy, it kind of tells you, you know,
link |
01:06:39.680
something to go do with data. But but, you know, our view is that it's and I think this is now,
link |
01:06:47.040
you know, quite widespread that it's, you know, despite the fact that those notions of
link |
01:06:52.640
anonymization kind of redacting and coarsening are the most widely adopted technical solutions
link |
01:07:00.720
for data privacy, they are like deeply fundamentally flawed. And so, you know, to your first question,
link |
01:07:07.120
what is differential privacy? Differential privacy seems to be a much, much better notion
link |
01:07:14.320
of privacy that kind of avoids a lot of the weaknesses of anonymization notions while
link |
01:07:22.000
while still letting us do useful stuff with data. What's anonymization of data?
link |
01:07:27.200
So by anonymization, I'm, you know, kind of referring to techniques like I have a database,
link |
01:07:32.640
the rows of that database are, let's say, individual people's medical records. Okay. And I
link |
01:07:41.440
I want to let people use that data. Maybe I want to let researchers access that data to
link |
01:07:46.480
build predictive models for some disease. But I'm worried that that will leak, you know,
link |
01:07:53.760
sensitive information about specific people's medical records. So anonymization broadly
link |
01:07:58.880
refers to the set of techniques where I say, like, okay, I'm first going to like, like, I'm
link |
01:08:03.040
going to delete the column with people's names. I'm going to not put, you know, so that would be
link |
01:08:08.240
like a redaction, right? I'm just redacting that information. I am going to take ages and I'm not
link |
01:08:15.200
going to like say your exact age, I'm going to say whether you're, you know, zero to 10, 10 to 20,
link |
01:08:20.720
20 to 30, I might put the first three digits of your zip code, but not the last two, etc, etc.
link |
01:08:27.200
And so the idea is that through some series of operations like this on the data, I anonymize it,
link |
01:08:31.680
you know, another term of art that's used is removing personally identifiable information.
link |
01:08:38.400
And, you know, this is basically the most common way of providing data privacy, but that's in a
link |
01:08:45.600
way that still lets people access the some variant form of the data. So at a slightly broader picture,
link |
01:08:52.320
as you talk about, what does anonymization mean when you have multiple database, like with a
link |
01:08:57.680
Netflix prize, when you can start combining stuff together. So this is exactly the problem
link |
01:09:03.680
with these notions, right? Is that notions of a don anonymization, removing personally
link |
01:09:08.320
identifiable information, the kind of fundamental conceptual flaw is that, you know, these definitions
link |
01:09:14.720
kind of pretend as if the data set in question is the only data set that exists in the world,
link |
01:09:20.720
or that ever will exist in the future. And of course, things like the Netflix prize and many,
link |
01:09:26.160
many other examples since the Netflix prize, I think that was one of the earliest ones, though,
link |
01:09:30.640
you know, you can reidentify people that were, you know, that were anonymized in the data set by
link |
01:09:37.360
taking that anonymized data set and combining it with other allegedly anonymized data sets and
link |
01:09:41.920
maybe publicly available information about you, you know, for people who don't know the Netflix
link |
01:09:46.320
prize was what was being publicly released as data. So the names from those rows were removed,
link |
01:09:53.200
but what was released is the preference or the ratings of what movies you like and you don't
link |
01:09:57.840
like. And from that combined with other things, I think forum posts and so on, you can start to figure
link |
01:10:03.760
out the names. That case, it was specifically the internet movie database, where lots of
link |
01:10:09.280
Netflix users publicly rate their movie, you know, their movie preferences. And so the anonymized data
link |
01:10:16.560
and Netflix when it's just this phenomenon, I think that we've all come to realize in the last
link |
01:10:23.680
decade or so is that just knowing a few apparently irrelevant innocuous things about you can often
link |
01:10:31.840
act as a fingerprint. Like if I know, you know, what rating you gave to these 10 movies and the
link |
01:10:38.480
date on which you entered these movies, this is almost like a fingerprint for you is in the sea
link |
01:10:43.120
of all Netflix users. There were just another paper on this in science or nature of about a
link |
01:10:48.160
month ago that, you know, kind of 18 attributes. I mean, my favorite example of this was actually
link |
01:10:54.400
a paper from several years ago now where it was shown that just from your likes on Facebook,
link |
01:11:01.600
just from the, you know, the things on which you clicked on the thumbs up button on the platform,
link |
01:11:07.360
not using any information, demographic information, nothing about who your friends are,
link |
01:11:12.480
just knowing the content that you had liked was enough to, you know, in the aggregate,
link |
01:11:19.360
accurately predict things like sexual orientation, drug and alcohol use, whether you were the child
link |
01:11:25.440
of divorced parents. So we live in this era where, you know, even the apparently irrelevant data that
link |
01:11:31.360
we offer about ourselves on public platforms and forums often unbeknownst to us, more or less acts as
link |
01:11:38.960
signature or, you know, fingerprint, and that if you can kind of, you know, do a join between that
link |
01:11:45.680
kind of data and allegedly anonymized data, you have real trouble. So is there hope for any kind
link |
01:11:52.240
of privacy in the world where a few likes can identify you? So there is differential privacy,
link |
01:11:59.920
right? So what is differential privacy? So differential privacy basically is a kind of
link |
01:12:04.240
alternate much stronger notion of privacy than these anonymization ideas. And it, you know,
link |
01:12:11.840
it's a technical definition, but like the spirit of it is we compare two alternate worlds. Okay, so
link |
01:12:20.640
let's suppose I'm a researcher and I want to do, you know, there's a database of medical records
link |
01:12:26.480
and one of them is yours. And I want to use that database of medical records to build a predictive
link |
01:12:32.480
model for some disease. So based on people's symptoms and test results and the like, I want to,
link |
01:12:38.560
you know, build a problem, you know, model predicting the probability that people have disease. So,
link |
01:12:42.240
you know, this is the type of scientific research that we would like to be allowed to continue.
link |
01:12:47.760
And in differential privacy, you act, ask a very particular counterfactual question.
link |
01:12:52.400
We basically compare two alternatives. One is when I do this, I build this model on the
link |
01:13:01.840
database of medical records, including your medical record. And the other one is where I do the same
link |
01:13:09.920
exercise with the same database with just your medical record removed. So basically, you know,
link |
01:13:16.640
it's two databases, one with n records in it, and one with n minus one records in it. The n minus
link |
01:13:23.680
one records are the same. And the only one that's missing in the second case is your medical record.
link |
01:13:29.600
So differential privacy basically says that any harms that might come to you from the analysis
link |
01:13:40.640
in which your data was included are essentially nearly identical to the harms that would have come
link |
01:13:47.600
to you if the same analysis had been done without your medical record included. So in other words,
link |
01:13:54.080
this doesn't say that bad things cannot happen to you as a result of data analysis. It just says
link |
01:13:59.600
that these bad things were going to happen to you already, even if your data wasn't included.
link |
01:14:05.200
And to give a very concrete example, right, you know, like we discussed at some length,
link |
01:14:11.760
the study that in the 50s that was done that established the link between smoking and lung
link |
01:14:18.560
cancer. And we make the point that like, well, if your data was used in that analysis,
link |
01:14:24.800
and the world kind of knew that you were a smoker because there was no stigma associated
link |
01:14:29.520
with smoking before that, those findings, real harm might have come to you as a result of that
link |
01:14:35.440
study that your data was included in. In particular, your insurer now might have a higher posterior
link |
01:14:41.040
belief that you might have lung cancer and raise your premium. So you've suffered economic damage.
link |
01:14:47.520
But the point is, is that if the same analysis has been done without with all the other n minus
link |
01:14:54.480
one medical records and just yours missing, the outcome would have been the same. Your data was
link |
01:15:00.240
an idiosyncratically crucial to establishing the link between smoking and lung cancer because the
link |
01:15:06.320
link between smoking and lung cancer is like a fact about the world that can be discovered with any
link |
01:15:11.840
sufficiently large database of medical records. But that's a very low value of harm. Yeah. So
link |
01:15:17.600
that's showing that very little harm is done. Great. But how, what is the mechanism of differential
link |
01:15:23.360
privacy? So that's the kind of beautiful statement of it. But what's the mechanism by which
link |
01:15:28.880
privacy is preserved? Yeah. So it's basically by adding noise to computations, right? So the
link |
01:15:34.560
basic idea is that every differentially private algorithm, first of all, or every good differentially
link |
01:15:41.280
private algorithm, every useful one is a probabilistic algorithm. So it doesn't, on a given input,
link |
01:15:48.080
if you gave the algorithm the same input multiple times, it would give different outputs each time
link |
01:15:53.280
from some distribution. And the way you achieve differential privacy algorithmically is by kind
link |
01:15:58.960
of carefully and tastefully adding noise to a computation in the right places. And, you know,
link |
01:16:05.440
to give a very concrete example, if I want to compute the average of a set of numbers, right,
link |
01:16:11.360
the non private way of doing that is to take those numbers and average them and release
link |
01:16:16.320
like a numerically precise value for the average. Okay. In differential privacy, you wouldn't do
link |
01:16:23.280
that. You would first compute that average to numerical precisions. And then you'd add some
link |
01:16:29.120
noise to it, right? You'd add some kind of zero mean, you know, Gaussian or exponential noise to
link |
01:16:35.600
it. So that the actual value you output, right, is not the exact mean, but it'll be close to the mean.
link |
01:16:42.240
But it'll be close. The noise that you add will sort of prove that nobody can kind of reverse
link |
01:16:48.160
engineer any particular value that went into the average. So noise, noise is the savior. How many
link |
01:16:55.760
algorithms can be aided by adding noise? Yeah, so I'm a relatively recent member of the differential
link |
01:17:05.360
privacy community. My coauthor Aaron Roth is, you know, really one of the founders of the field
link |
01:17:10.560
and has done a great deal of work. And I've learned a tremendous amount working with him on it.
link |
01:17:14.880
It's a pretty grown up field already. Yeah, but now it's pretty mature. But I must admit, the first
link |
01:17:18.880
time I saw the definition of differential privacy, my reaction was like, well, that is a clever
link |
01:17:23.440
definition. And it's really making very strong promises. And my, you know, you know, I first
link |
01:17:29.920
saw the definition in much earlier days. And my first reaction was like, well, my worry about
link |
01:17:34.880
this definition would be that it's a great definition of privacy, but that it'll be so
link |
01:17:38.880
restrictive that we won't really be able to use it. Like, you know, we won't be able to do compute
link |
01:17:43.920
many things in a differentially private way. So that's one of the great successes of the field,
link |
01:17:49.360
I think, is in showing that the opposite is true. And that, you know, most things that we know how
link |
01:17:56.000
to compute absent any privacy considerations can be computed in a differentially private way. So
link |
01:18:02.400
for example, pretty much all of statistics and machine learning can be done differentially
link |
01:18:07.280
privately. So pick your favorite machine learning algorithm, back propagation and neural networks,
link |
01:18:13.840
you know, cart for decision trees, support vector machines, boosting, you name it,
link |
01:18:20.000
as well as classic hypothesis testing and the like and statistics. None of those algorithms
link |
01:18:26.320
are differentially private in their original form. All of them have modifications that add
link |
01:18:32.960
noise to the computation in different places in different ways that achieve differential privacy.
link |
01:18:38.880
So this really means that to the extent that, you know, we've become a, you know, a scientific
link |
01:18:45.280
community very dependent on the use of machine learning and statistical modeling and data analysis,
link |
01:18:51.920
we really do have a path to kind of provide privacy guarantees to those methods. And so we
link |
01:18:59.280
can still, you know, enjoy the benefits of kind of the data science era while providing, you know,
link |
01:19:06.960
rather robust privacy guarantees to individuals. So perhaps a slightly crazy question, but if we
link |
01:19:13.760
take that, the idea is a differential privacy and take it to the nature of truth that's being
link |
01:19:18.960
explored currently. So what's your most favorite and least favorite food?
link |
01:19:24.160
Hmm. I'm not a real foodie. So I'm a big fan of spaghetti.
link |
01:19:29.600
Spaghetti? What do you really don't like?
link |
01:19:35.600
I really don't like cauliflower.
link |
01:19:37.920
Well, I love cauliflower. Okay. But is one way to protect your preference for spaghetti by
link |
01:19:44.320
having an information campaign, bloggers and so on, of bots saying that you like cauliflower?
link |
01:19:50.400
So like this kind of the same kind of noise ideas. I mean, if you think of in our politics today,
link |
01:19:57.040
there's this idea of Russia hacking our elections. What's meant there, I believe,
link |
01:20:03.600
is bots spreading different kinds of information. Is that a kind of privacy or is that too much
link |
01:20:09.440
of a stretch? No, it's not a stretch. I've not seen those ideas, you know, that is not a technique
link |
01:20:17.520
that to my knowledge will provide differential privacy. But to give an example, like one very
link |
01:20:23.200
specific example about what you're discussing is there was a very interesting project at NYU,
link |
01:20:29.840
I think led by Helen Nissenbaum there, in which they basically built a browser plugin that
link |
01:20:38.320
tried to essentially obfuscate your Google searches. So to the extent that you're worried
link |
01:20:43.280
that Google is using your searches to build, you know, predictive models about you to decide what
link |
01:20:49.520
ads to show you, which they might very reasonably want to do. But if you object to that, they built
link |
01:20:54.960
this widget you could plug in. And basically, whenever you put an aquarium to Google, it would
link |
01:21:00.000
send that query to Google. But in the background, all of the time from your browser, it would just
link |
01:21:05.200
be sending this torrent of irrelevant queries to the search engine. So, you know, it's like a weed
link |
01:21:12.960
and chaff thing. So, you know, out of every 1000 queries, let's say that Google was receiving
link |
01:21:18.560
from your browser, one of them was one that you put in, but the other 999 were not. Okay, so it's
link |
01:21:24.160
the same kind of idea, kind of, you know, privacy by obfuscation. So I think that's an interesting
link |
01:21:30.800
idea. Doesn't give you differential privacy. It's also, I was actually talking to somebody at one
link |
01:21:37.520
of the large tech companies recently about the fact that, you know, just this kind of thing that
link |
01:21:43.280
there are sometimes when the response to my data needs to be very specific to my data, right?
link |
01:21:51.520
Like I type mountain biking into Google, I want results on mountain biking, and I really want
link |
01:21:57.040
Google to know that I typed in mountain biking. I don't want noise added to that. And so I think
link |
01:22:02.960
there's sort of maybe even interesting technical questions around notions of privacy that are
link |
01:22:07.040
appropriate where, you know, it's not that my data is part of some aggregate like medical records and
link |
01:22:12.720
that we're trying to discover important correlations and facts about the world at large, but rather,
link |
01:22:18.640
you know, there's a service that I really want to, you know, pay attention to my specific data,
link |
01:22:23.920
yet I still want some kind of privacy guarantee. And I think these kind of obfuscation ideas are
link |
01:22:28.560
sort of one way of getting at that, but maybe there are others as well.
link |
01:22:31.920
So where do you think we'll land in this algorithm driven society in terms of privacy? So
link |
01:22:37.520
sort of China, like Kaifu Lee describes, you know, it's collecting a lot of data on its citizens,
link |
01:22:45.920
but in the best form, it's actually able to provide a lot of sort of protect human rights and
link |
01:22:52.880
provide a lot of amazing services. And it's worse forms that can violate those human rights and
link |
01:22:59.680
limit services. So where do you think we'll land? So algorithms are powerful when they use data.
link |
01:23:08.160
So as a society, do you think we'll give over more data? Is it possible to protect the privacy of
link |
01:23:14.880
that data? So I'm optimistic about the possibility of, you know, balancing the desire for individual
link |
01:23:24.480
privacy and individual control of privacy with kind of societally and commercially beneficial
link |
01:23:32.400
uses of data, not unrelated to differential privacy or suggestions that say like, well,
link |
01:23:37.920
individuals should have control of their data. They should be able to limit the uses of that data.
link |
01:23:43.440
They should even, you know, there's, there's, you know, fledgling discussions going on in
link |
01:23:47.760
research circles about allowing people selective use of their data and being compensated for it.
link |
01:23:54.480
And then you get to sort of very interesting economic questions like pricing, right?
link |
01:23:59.200
And one interesting idea is that maybe differential privacy would also, you know,
link |
01:24:04.080
be a conceptual framework in which you could talk about the relative value of different people's
link |
01:24:08.800
data, like, you know, to demystify this a little bit. If I'm trying to predict, build a predictive
link |
01:24:13.680
model for some rare disease, and I'm trying to use, I'm going to use machine learning to do it,
link |
01:24:19.520
it's easy to get negative examples because the disease is rare, right? But I really want
link |
01:24:24.560
to have lots of people with the disease in my data set. Okay. But, but, and so somehow those
link |
01:24:31.600
people's data with respect to this application is much more valuable to me than just like the
link |
01:24:36.560
background population. And so maybe they should be compensated more for it. And so, you know,
link |
01:24:43.520
I think these are kind of very, very fledgling conceptual questions that maybe will have kind
link |
01:24:50.160
of technical thought on them sometime in the coming years. But, but I do think we'll,
link |
01:24:54.480
you know, to kind of get more directly answer your question, I think I'm optimistic at this
link |
01:24:59.120
point from what I've seen that we will land at some, you know, better compromise than we're at
link |
01:25:04.960
right now, where again, you know, privacy guarantees are few far between and weak,
link |
01:25:11.440
and users have very, very little control. And I'm optimistic that we'll land in something that,
link |
01:25:17.840
you know, provides better privacy overall and more individual control of data and privacy. But,
link |
01:25:22.800
you know, I think to get there, it's again, just like fairness, it's not going to be enough to
link |
01:25:27.840
propose algorithmic solutions. There's going to have to be a whole kind of regulatory legal process
link |
01:25:32.880
that prods companies and other parties to kind of adopt solutions.
link |
01:25:38.160
And I think you've mentioned the word control a lot. And I think giving people control,
link |
01:25:42.720
that's something that people don't quite have in a lot of these algorithms. And that's a really
link |
01:25:48.160
interesting idea of giving them control. Some of that is actually literally an interface design
link |
01:25:53.920
question, sort of just enabling, because I think it's good for everybody to give users control.
link |
01:26:00.240
It's not, it's not, it's almost not a trade off, except that you have to hire people that are good
link |
01:26:05.600
at interface design. Yeah, I mean, the other thing that has to be said, right, is that, you know,
link |
01:26:11.840
it's a cliche, but, you know, we, as the users of many systems, platforms and apps,
link |
01:26:19.040
you know, we are the product, we are not the customer. The customer are advertisers, and
link |
01:26:24.960
our data is the product. Okay. So it's one thing to kind of suggest more individual control of
link |
01:26:31.280
data and privacy and uses. But this, you know, if this happens in sufficient degree, it will
link |
01:26:38.960
upend the entire economic model that has supported the internet to date. And so some other economic
link |
01:26:45.840
model will have to be, you know, we'll have to replace it. So the idea of Mark, as you mentioned,
link |
01:26:51.440
and by exposing the economic model to the people, they will then become a market.
link |
01:26:57.520
They could be participants in it. Participants in it.
link |
01:26:59.840
And, and, you know, this isn't, you know, this is not a weird idea, right? Because
link |
01:27:03.760
there are markets for data already. It's just that consumers are not participants in them.
link |
01:27:08.400
There's like, you know, there's sort of, you know, publishers and content providers on one side
link |
01:27:13.120
that have inventory and then they're advertised on others. And, you know, you know, Google and
link |
01:27:17.840
Facebook are running, you know, they're pretty much their entire revenue stream is by running
link |
01:27:23.920
two sided markets between those parties, right? And so it's not a crazy idea that there would be
link |
01:27:29.600
like a three sided market or that, you know, that on one side of the market or the other,
link |
01:27:34.320
we would have proxies representing our interest. It's not, you know, it's not a crazy idea,
link |
01:27:39.120
but it would, it's not a crazy technical idea, but it would have
link |
01:27:43.920
pretty extreme economic consequences. Speaking of markets, a lot of fascinating
link |
01:27:51.920
aspects of this world arise not from individual humans, but from the interaction of human beings.
link |
01:27:59.680
You've done a lot of work in game theory. First, can you say, what is game theory and how does
link |
01:28:05.760
help us model and study? Yeah, game theory, of course, let us give credit where it's due.
link |
01:28:10.880
It comes from the economist first and foremost. But as I mentioned before, like, you know,
link |
01:28:16.720
computer scientists never hesitate to wander into other people's turf. And so there is now this
link |
01:28:23.280
20 year old field called algorithmic game theory. But, you know, game theory, first and foremost,
link |
01:28:30.080
is a mathematical framework for reasoning about collective outcomes in systems of interacting
link |
01:28:37.760
individuals. You know, so you need at least two people to get started in game theory. And
link |
01:28:45.200
many people are probably familiar with prisoner's dilemma as kind of a classic example of game
link |
01:28:50.160
theory and a classic example where everybody looking out for their own individual interests
link |
01:28:56.640
leads to a collective outcome that's kind of worse for everybody than what might be possible if they
link |
01:29:03.040
cooperate it, for example. But cooperation is not an equilibrium in prisoner's dilemma.
link |
01:29:09.520
And so my work and the field of algorithmic game theory more generally in these areas
link |
01:29:16.000
kind of looks at settings in which the number of actors is potentially extraordinarily large
link |
01:29:24.640
and their incentives might be quite complicated and kind of hard to model directly. But you still
link |
01:29:31.360
want kind of algorithmic ways of kind of predicting what will happen or influencing what will happen
link |
01:29:36.720
in the design of platforms. So what to you is the most beautiful idea that you've encountered in
link |
01:29:44.800
game theory? There's a lot of them. I'm a big fan of the field. I mean, you know, I mean technical
link |
01:29:52.480
answers to that, of course, would include Nash's work just establishing that, you know, there's
link |
01:29:59.120
a competitive equilibrium under very, very general circumstances, which in many ways kind of put the
link |
01:30:05.200
field on a firm conceptual footing because if you don't have equilibrium, it's kind of hard to ever
link |
01:30:11.920
reason about what might happen since, you know, there's just no stability.
link |
01:30:15.920
So just the idea of the stability can emerge when there's multiple.
link |
01:30:19.680
Or not that it will necessarily emerge just that it's possible, right? Like the existence of
link |
01:30:24.560
equilibrium doesn't mean that sort of natural, iterative behavior will necessarily lead to it.
link |
01:30:30.400
In the real world, yes.
link |
01:30:31.840
Maybe answering a slightly less personally than you asked the question. I think within the
link |
01:30:36.000
field of algorithmic game theory, perhaps the single most important kind of technical
link |
01:30:43.680
contribution that's been made is the realization between close connections between
link |
01:30:49.120
machine learning and game theory and in particular between game theory and
link |
01:30:52.800
the branch of machine learning that's known as no regret learning.
link |
01:30:56.240
And this sort of provides a very general framework in which a bunch of players interacting
link |
01:31:04.000
in a game or a system, each one kind of doing something that's in their self interest will
link |
01:31:09.840
actually kind of reach an equilibrium and actually reach an equilibrium in a pretty,
link |
01:31:16.320
you know, a rather, you know, short amount of steps.
link |
01:31:20.240
So you kind of mentioned acting greedily can somehow end up pretty good for everybody.
link |
01:31:29.680
Or pretty bad.
link |
01:31:31.120
Or pretty bad. It will end up stable.
link |
01:31:34.240
Yeah, right. And, you know, stability or equilibrium by itself is not necessarily either a good thing
link |
01:31:41.920
or a bad thing.
link |
01:31:42.960
So what's the connection between machine learning and the ideas?
link |
01:31:45.600
Well, I think we kind of talked about these ideas already in kind of a non technical way,
link |
01:31:50.880
which is maybe the more interesting way of understanding them first, which is, you know,
link |
01:31:57.040
we have many systems, platforms and apps these days that work really hard to use our data and
link |
01:32:04.880
the data of everybody else on the platform to selfishly optimize on behalf of each user.
link |
01:32:11.840
Okay. So, you know, let me let me give I think the cleanest example, which is just driving apps,
link |
01:32:17.920
navigation apps like, you know, Google Maps and Waze, where, you know, miraculously compared to
link |
01:32:24.080
when I was growing up at least, you know, the objective would be the same when you wanted
link |
01:32:28.640
to drive from point A to point B, spend the least time driving, not necessarily minimize the distance,
link |
01:32:34.320
but minimize the time, right. And when I was growing up, like the only resources you had to
link |
01:32:39.280
do that were like maps in the car, which literally just told you what roads were available. And then
link |
01:32:45.440
you might have like half hourly traffic reports, just about the major freeways, but not about
link |
01:32:51.040
side roads. So you were pretty much on your own. And now we've got these apps, you pull it out and
link |
01:32:56.400
you say, I want to go from point A to point B. And in response kind of to what everybody else is
link |
01:33:01.680
doing, if you like, what all the other players in this game are doing right now, here's the, you
link |
01:33:07.360
know, the, the route that minimizes your driving time. So it is really kind of computing a selfish
link |
01:33:13.520
best response for each of us in response to what all of the rest of us are doing at any given moment.
link |
01:33:20.000
And so, you know, I think it's quite fair to think of these apps as driving or nudging us all
link |
01:33:26.880
towards the competitive or Nash equilibrium of that game. Now you might ask, like, well,
link |
01:33:33.840
that sounds great. Why is that a bad thing? Well, you know, it's, it's known both in theory and
link |
01:33:42.000
with some limited studies from actual like traffic data that all of us being in this competitive
link |
01:33:50.640
equilibrium might cause our collective driving time to be higher, maybe significantly higher
link |
01:33:56.800
than it would be under other solutions. And then you have to talk about what those other
link |
01:34:01.040
solutions might be and what, what the algorithms to implement them are, which we do discuss in the
link |
01:34:06.320
kind of game theory chapter of the book. But, but similarly, you know, on social media platforms
link |
01:34:13.280
or on Amazon, you know, all these algorithms that are essentially trying to optimize our behalf,
link |
01:34:20.000
they're driving us in a colloquial sense towards some kind of competitive equilibrium.
link |
01:34:24.960
And, you know, one of the most important lessons of game theory is that just because
link |
01:34:28.400
we're at equilibrium doesn't mean that there's not a solution in which some or maybe even all of us
link |
01:34:33.520
might be better off. And then the connection to machine learning, of course, is that in all these
link |
01:34:38.480
platforms I've mentioned, the optimization that they're doing on our behalf is driven by machine
link |
01:34:43.680
learning, you know, like predicting where the traffic will be, predicting what products I'm
link |
01:34:47.840
going to like, predicting what would make me happy in my news feed. Now, in terms of the stability
link |
01:34:53.520
and the promise of that, I have to ask just out of curiosity, how stable are these mechanisms
link |
01:34:58.800
that you game theory is just the economists came up with. And we all know that economists don't
link |
01:35:04.080
live in the real world. Just kidding. So what's, do you think when we look at the fact that we
link |
01:35:11.760
haven't blown ourselves up from the game theoretic concept of mutually shared destruction,
link |
01:35:18.400
what are the odds that we destroy ourselves with nuclear weapons as one example of a stable
link |
01:35:25.680
game theoretic system? Just to prime your viewers a little bit, I mean, I think you're referring
link |
01:35:31.520
to the fact that game theory was taken quite seriously back in the 60s as a tool for reasoning
link |
01:35:37.920
about kind of Soviet US nuclear armament, disarmative detente, things like that.
link |
01:35:44.400
I'll be honest, as huge of a fan as I am of game theory and its kind of rich history,
link |
01:35:51.520
it still surprises me that, you know, you had people at the Rand Corporation back in those days
link |
01:35:57.680
kind of drawing up, you know, two by two tables and one, the role player is, you know, the US and
link |
01:36:02.800
the Colin players, Russia, and that they were taking seriously, you know, I'm sure if I was there,
link |
01:36:08.720
maybe it wouldn't have seemed as naive as it does at the time. It seems to have worked,
link |
01:36:13.680
which is why it seems naive. Well, we're still here. We're still here in that sense.
link |
01:36:17.840
Yeah. Even though I kind of laugh at those efforts, they were more sensible then than
link |
01:36:22.560
they would be now, right? Because there were sort of only two nuclear powers at the time,
link |
01:36:26.560
and you didn't have to worry about deterring new entrants and who was developing the capacity.
link |
01:36:32.400
And so we have many, you know, we have this, it's definitely a game with more players now and more
link |
01:36:38.640
potential entrants. I'm not in general somebody who advocates using kind of simple mathematical
link |
01:36:45.280
models when the stakes are as high as things like that and the complexities are very political and
link |
01:36:51.840
social, but we are still here. So you've worn many hats, one of which, the one that first
link |
01:36:59.280
caused me to become a big fan of your work many years ago is algorithmic trading. So I have to
link |
01:37:05.680
just ask a question about this because you have so much fascinating work there. In the 21st century,
link |
01:37:10.720
what role do you think algorithms have in the space of trading, investment in the financial
link |
01:37:17.520
sector? Yeah. It's a good question. I mean, in the time I've spent on Wall Street and in finance,
link |
01:37:26.880
you know, I've seen a clear progression. And I think it's a progression that kind of models
link |
01:37:31.120
the use of algorithms and automation more generally in society, which is, you know,
link |
01:37:37.840
the things that kind of get taken over by the algos first are sort of the things that computers are
link |
01:37:44.240
obviously better at than people, right? So, you know, so first of all, there needed to be this
link |
01:37:50.320
era of automation, right, where just, you know, financial exchanges became largely electronic,
link |
01:37:55.280
which then enabled the possibility of, you know, trading becoming more algorithmic because once,
link |
01:38:02.240
you know, that exchanges are electronic, an algorithm can submit an order through an API just
link |
01:38:07.360
as well as a human can do at a monitor. You can do it really quickly. You can read all the data.
link |
01:38:10.960
So yeah. And so, you know, I think the places where algorithmic trading have had the greatest
link |
01:38:18.160
inroads and had the first inroads were in kind of execution problems, kind of optimized execution
link |
01:38:24.160
problems. So what I mean by that is at a large brokerage firm, for example, one of the lines of
link |
01:38:30.000
business might be on behalf of large institutional clients taking, you know, what we might consider
link |
01:38:36.400
difficult trade. So it's not like a mom and pop investor saying, I want to buy 100 shares of
link |
01:38:40.720
Microsoft. It's a large hedge fund saying, you know, I want to buy a very, very large stake in
link |
01:38:47.200
Apple. And I want to do it over the span of a day. And it's such a large volume that if you're not
link |
01:38:53.280
clever about how you break that trade up, not just over time, but over perhaps multiple different
link |
01:38:58.640
electronic exchanges that all let you trade Apple on their platform, you know, you will, you will
link |
01:39:03.520
move, you'll push prices around in a way that hurts your, your execution. So you know, this is
link |
01:39:09.520
the kind of, you know, this is an optimization problem. This is a control problem. Right. And so
link |
01:39:16.240
machines are better. We know how to design algorithms, you know, that are better at that
link |
01:39:20.880
kind of thing than a person is going to be able to do because we can take volumes of historical
link |
01:39:26.000
and real time data to kind of optimize the schedule with which we trade. And, you know,
link |
01:39:31.040
similarly high frequency trading, you know, which is closely related, but not the same as
link |
01:39:36.400
optimized execution, where you're just trying to spot very, very temporary, you know, mispricings
link |
01:39:44.000
between exchanges or within an asset itself, or just predict directional movement of a stock
link |
01:39:50.160
because of the kind of very, very low level granular buying and selling data in the, in the
link |
01:39:56.880
exchange. Machines are good at this kind of stuff. It's kind of like the mechanics of trading.
link |
01:40:02.080
What about the, can machines do long terms of prediction? Yeah. So I think we are in an era
link |
01:40:09.760
where, you know, clearly there have been some very successful, you know,
link |
01:40:13.360
you know, quant hedge funds that are, you know, in what we would traditionally call, you know,
link |
01:40:19.760
still in the stat Arb regime, like so, you know, what's that stat Arb referring to statistical
link |
01:40:25.040
arbitrage. But, but for the purposes of this conversation, what it really means is making
link |
01:40:29.760
directional predictions in asset price movement or returns, your prediction about that directional
link |
01:40:37.520
movement is good for, you know, you have a view that it's valid for some period of time between
link |
01:40:44.480
a few seconds and a few days. And that's the amount of time that you're going to kind of get
link |
01:40:49.440
into the position, hold it and then hopefully be right about the directional movement and, you
link |
01:40:53.920
know, buy low and sell high as the cliche goes. So that is a, you know, kind of a sweet spot,
link |
01:41:00.800
I think, for quant trading and investing right now and has been for some time.
link |
01:41:06.000
When you really get to kind of more Warren Buffett style time scales, right? Like, you know,
link |
01:41:12.160
my cartoon of Warren Buffett is that, you know, Warren Buffett sits and thinks what the long
link |
01:41:16.800
term value of Apple really should be. And he doesn't even look at what Apple is doing today.
link |
01:41:22.480
He just decides, you know, you know, I think that this was what its long term value is and
link |
01:41:27.520
it's far from that right now. And so I'm going to buy some Apple or, you know, short some Apple.
link |
01:41:32.400
And I'm going to, I'm going to sit on that for 10 or 20 years. Okay. So when you're at that kind of
link |
01:41:39.840
timescale or even more than just a few days, all kinds of other sources of risk and information,
link |
01:41:48.880
you know, so now you're talking about holding things through recessions and economic cycles.
link |
01:41:54.000
Wars can break out. So there you have to understand human nature at a level.
link |
01:41:58.800
Yeah. And you need to just be able to ingest many, many more sources of data that are on wildly
link |
01:42:04.480
different timescales, right? So if I'm an HFT, my high frequency trader, like, I don't, I don't,
link |
01:42:11.280
I really, my main source of data is just the data from the exchanges themselves about the
link |
01:42:16.320
activity in the exchanges, right? And maybe I need to pay, you know, I need to keep an eye on the
link |
01:42:21.520
news, right? Because, you know, that can suddenly cause sudden, you know, the, you know, CEO gets
link |
01:42:26.960
caught in a scandal or, you know, gets run over by a bus or something that can cause very sudden
link |
01:42:31.600
changes. But, you know, I don't need to understand economic cycles. I don't need to understand
link |
01:42:36.960
recessions. I don't need to worry about the political situation or war breaking out in this
link |
01:42:42.160
part of the world, because, you know, all you need to know is as long as that's not going to happen
link |
01:42:46.960
in the next 500 milliseconds, then, you know, my model is good. When you get to these longer
link |
01:42:53.280
time scales, you really have to worry about that kind of stuff. And people in the machine learning
link |
01:42:56.800
community are starting to think about this. We held a, we jointly sponsored a workshop at Penn
link |
01:43:04.000
with the Federal Reserve Bank of Philadelphia a little more than a year ago on, you know, I think
link |
01:43:09.120
the title is something like machine learning for macroeconomic prediction, you know, macroeconomic
link |
01:43:15.120
referring specifically to these longer time scales. And, you know, it was an interesting
link |
01:43:20.000
conference, but it, you know, it left me with greater confidence that you have a long way to
link |
01:43:28.320
go to, you know, and so I think that people that, you know, in the grand scheme of things, you know,
link |
01:43:33.440
if somebody asks me like, well, whose job on Wall Street is safe from the bots, I think people
link |
01:43:39.200
that are at that longer, you know, time scale and have that appetite for all the risks involved in
link |
01:43:44.400
long term investing and that really need kind of not just algorithms that can optimize from data,
link |
01:43:50.640
but they need views on stuff. They need views on the political landscape, economic cycles and the
link |
01:43:56.640
like. And I think, you know, they're, they're, they're pretty safe for a while, as far as I can
link |
01:44:02.240
tell. So Warren Buffett's job is safe. Yeah, I'm not seeing, you know, a robo Warren Buffett
link |
01:44:07.440
any time soon. Give him comfort. Last question. If you could go back to if there's a day in your
link |
01:44:16.480
life, you could relive because it made you truly happy. Maybe you outside the family.
link |
01:44:23.600
Yeah, otherwise, you know, what, what day would it be? What can you look back? You remember just
link |
01:44:31.280
being profoundly transformed in some way or blissful?
link |
01:44:40.160
I'll answer a slightly different question, which is like, what's this a day in my, my life or my
link |
01:44:44.960
career that was kind of a watershed moment? I went straight from undergrad to doctoral studies. And,
link |
01:44:53.120
you know, that's not at all atypical. And I'm also from an academic family, like my, my dad was a
link |
01:44:58.400
professor, my uncle on his side as a professor, both my grandfathers were professors. All kinds
link |
01:45:03.680
of majors to philosophy. So yeah, they're kind of all over the map. Yeah. And I was a grad student
link |
01:45:10.160
here just up the river at Harvard and came to study with Les Valley, which was a wonderful
link |
01:45:14.400
experience. But you know, I remember my first year of graduate school, I was generally pretty
link |
01:45:19.840
unhappy. And I was unhappy because, you know, at Berkeley as an undergraduate, you know, yeah,
link |
01:45:25.200
I studied a lot of math and computer science. But it was a huge school, first of all. And I took a
link |
01:45:29.920
lot of other courses, as we discussed, I started as an English major and took history courses and
link |
01:45:34.640
art history classes, and had friends, you know, that did all kinds of different things. And,
link |
01:45:40.160
you know, Harvard's a much smaller institution than Berkeley. And its computer science department,
link |
01:45:44.720
especially at that time, was, was a much smaller place than it is now. And I suddenly just felt
link |
01:45:50.080
very, you know, like, I'd gone from this very big world to this highly specialized world.
link |
01:45:56.640
And now all of the classes I was taking were computer science classes. And I was only in
link |
01:46:01.280
classes with math and computer science people. And so I was, you know, I thought often in that
link |
01:46:08.160
first year of grad school about whether I really wanted to stick with it or not. And, you know,
link |
01:46:12.800
I thought like, oh, I could, you know, stop with a master's, I could go back to the Bay Area and
link |
01:46:18.000
to California. And, you know, this was in one of the early periods where there was, you know,
link |
01:46:22.160
like, you could definitely get a relatively good job paying job at one of the, one of the tech
link |
01:46:27.600
companies back, you know, that were the big tech companies back then. And so I distinctly remember
link |
01:46:32.640
like kind of a late spring day when I was kind of, you know, sitting in Boston Common and kind of
link |
01:46:38.240
really just kind of chewing over what I wanted to do with my life. And I realized like, okay,
link |
01:46:43.040
you know, and I think this is where my academic background helped me a great deal. I sort of
link |
01:46:46.240
realized, you know, yeah, you're not having a great time right now. This feels really narrowing,
link |
01:46:51.440
but you know that you're here for research eventually and to do something original and to
link |
01:46:56.800
try to, you know, carve out a career where you kind of, you know, choose what you want to think
link |
01:47:02.800
about, you know, and have a great deal of independence. And so, you know, at that point,
link |
01:47:07.920
I really didn't have any real research experience yet. I mean, it was trying to think about some
link |
01:47:12.240
problems with very little success, but I knew that like I hadn't really tried to do the thing
link |
01:47:19.840
that I knew I'd come to do. And so I thought, you know, I'm gonna stick through it for the summer
link |
01:47:26.880
and, you know, and that was very formative because I went from kind of contemplating quitting to,
link |
01:47:35.040
you know, a year later, it being very clear to me I was going to finish because I still had
link |
01:47:39.680
a ways to go, but I kind of started doing research. It was going well. It was really
link |
01:47:44.960
interesting and it was sort of a complete transformation. You know, it's just that transition
link |
01:47:49.280
that I think every doctoral student makes at some point, which is to sort of go from being like a
link |
01:47:55.840
student of what's been done before to doing, you know, your own thing and figure out what makes
link |
01:48:02.960
you interested in what your strengths and weaknesses are as a researcher. And once, you
link |
01:48:07.280
know, I kind of made that decision on that particular day at that particular moment in
link |
01:48:11.760
Boston Common. You know, I'm glad I made that decision. And also just accepting the painful
link |
01:48:17.920
nature of that journey. Yeah, exactly. Exactly. And in that moment said, I'm gonna stick it out.
link |
01:48:24.240
Yeah, I'm gonna stick around for a while. Well, Michael, I've looked up to you work for a long
link |
01:48:29.440
time. It's really nice to talk to you. Thank you so much for doing it. It's great to get back in touch
link |
01:48:32.000
with you too and see how great you're doing as well. Thank you. Thanks a lot. Appreciate it.