Moritz Hardt: So I’m a com­put­er sci­en­tist, and I called this response Occupy Algorithms: Will Algorithms Serve the 99%.” And more than sort of a dis­cus­sion of what’s been said so far this is a kind of research pro­pos­al of what I would like to see hap­pen­ing at the inter­sec­tion of CS and this audience. 

So, what Frank is get­ting at in my opin­ion is sort of the ques­tion have albums in finance failed to deliv­er on the promise of greater effi­cien­cy, trans­paren­cy, and fair­ness? And so what I want to address in this response is gee, what have com­put­er sci­en­tist done about it, right, if any­thing? And I picked out three works that I think are very close­ly relat­ed to what Frank talked about. 

The first one’s about ensur­ing fair­ness in clas­si­fi­ca­tion as a press­ing theme in Frank’s arti­cle. The sec­ond one is on thwart­ing gam­ing and manip­u­la­tion. And gam­ing came up I think so far in every talk and every response, so I think this is also relat­ed. And final­ly I want to talk about the role of com­pu­ta­tion­al com­plex­i­ty in finan­cial mar­kets and finan­cial prod­ucts, and the dif­fi­cul­ty that this intro­duces in the reg­u­la­to­ry process and the pric­ing process. And I hope you all remem­ber what com­pu­ta­tion­al com­plex­i­ty is from Bob’s talk and what the word intractable means. 

So I should say this is not an exhaus­tive list. There are many more fas­ten­ing papers in the CS lit­er­a­ture. These are just some that I thought were real­ly relat­ed. And yeah, each is just just a two-minute teas­er on these papers. I just have a few min­utes, so I prob­a­bly won’t make any true state­ments. But you can look at the papers if you actu­al­ly want to see the truth. 

So, the first paper I want to men­tion, we called it Fairness Through Awareness.” This was joint work with Cynthia Dwork, Omer Reingold, Toni Pitassi, and Rich Zemel from about two years ago. And the prob­lem that we’re get­ting at from sort of a math­e­mat­i­cal per­spec­tive was how do we pre­vent a clas­si­fi­er or clas­si­fy­ing algo­rithm from dis­crim­i­nat­ing based on pro­tect­ed attrib­ut­es such as you know, race, or reli­gion, or gen­der? And the run­ning exam­ple in Frank’s arti­cle would be the cred­it loan appli­ca­tion where you’re con­cerned about for exam­ple steer­ing minori­ties into less-advantageous cred­it card offers. So this was also a moti­vat­ing exam­ple for us, but also the tar­get­ed adver­tis­ing exam­ple that we saw yes­ter­day is real­ly good here, and fitting.

So, let me just sort of recall of what an insuf­fi­cient solu­tion con­cept is, just because a lot of com­pa­nies will actu­al­ly tell you that that’s what they’re doing. And I call this blind­ness to the pro­tect­ed attribute. So, peo­ple will tell you, Oh, we don’t even look at the race attribute. How could our algo­rithm be dis­crim­i­nat­ing if race does­n’t show up any­where in the data and we nev­er looked at that attribute?” And as every machine learn­ing per­son knows, there are lots of redun­dant encod­ings of race in the data. You don’t need to see the attribute to be able to reli­ably pre­dict it. So this does­n’t work giv­en our under­stand­ing of machine learn­ing. It’s just not a solution. 

So what we end­ed up doing is we came up with a new notion which we called individual-based fair­ness. And it’s a for­mal­iza­tion of this idea here, which says that any two indi­vid­u­als who are sim­i­lar with respect to the spe­cif­ic clas­si­fi­ca­tion task that you’re inter­est­ed in should receive sim­i­lar out­comes or oth­er sim­i­lar dis­tri­b­u­tion over out­comes. And this is just an infor­mal state­ment of what our for­mal notion says. But what’s impor­tant here, what should strike you as impor­tant, is that it requires the def­i­n­i­tion of a task-specific met­ric or task-specific sim­i­lar­i­ty mea­sure between indi­vid­u­als that allows you to com­pare individuals. 

So this is real­ly a cen­tral part of this def­i­n­i­tion, and so I want to empha­size this a lit­tle bit more. So what’s this met­ric? We think of it as exter­nal­ly imposed and open to the pub­lic. It’s some­thing that peo­ple thought about, and as such it pro­vides trans­paren­cy. This is some­thing that peo­ple who’re inter­est­ed in the issue of fair­ness could dis­cuss and influ­ence. So this is some­thing that Frank likes because it gives you a way to reg­u­late the clas­si­fi­ca­tion process. 

On the oth­er hand, the met­ric does not dic­tate clas­si­fi­ca­tion. There could be many clas­si­fiers, many very dif­fer­ent clas­si­fiers, that all con­sis­tent with the same met­ric. So in par­tic­u­lar it still allows the bank to opti­mize their secret, pro­pri­etary strat­e­gy. And this is some­thing that Tal real­ly thinks is impor­tant for com­pe­ti­tion. So this met­ric gives you a way to go between these stand­points and get a tradeoff. 

But one open prob­lem that we always face and to me as a com­put­er sci­en­tist I can’t ful­ly solve is, in any con­crete set­ting who gets to define the met­ric, and how? And this is real­ly some­thing very inter­dis­ci­pli­nary where we need to rely on social insight and peo­ple from var­i­ous dis­ci­plines here in this room. And that’s sort of an open prob­lem I have for you. So this is where my work as a com­put­er sci­en­tist sort of ends. If you give me some­thing like this, if you can think about what it might be in var­i­ous set­tings that you’re inter­est­ed in, I can give you a use­ful algo­rith­mic framework. 

So just as a sort of a car­toon of this thing, there is the set of indi­vid­u­als. There could be two indi­vid­u­als, they are relat­ed by the met­ric, so they’re at a cer­tain dis­tance. And then the frame­work would make sure that the dis­tri­b­u­tion over out­comes that they receive is sim­i­lar to the extent that they’re sim­i­lar with respect to the met­ric. This is just a car­toon. You can look in the paper for the actu­al definition. 

So I want to move on, and I want to get to the sec­ond thing I men­tioned. And this is about gam­ing clas­si­fiers and how you might be able to pre­vent it on an algo­rith­mic lev­el. And one thing that I real­ly like is this wis­dom from finan­cial pol­i­cy­mak­ing called Goodhart’s Law. And this start­ed my inter­est in this ques­tion a long time ago. And Goodhart’s Law says that any observed sta­tis­ti­cal reg­u­lar­i­ty will tend to col­lapse once pres­sure is placed upon it for con­trol pur­pos­es.” And if you trans­late this into machine learn­ing jar­gon, what it says is if you have a clas­si­fi­er that relies on some sta­tis­ti­cal pat­tern, and you expose the clas­si­fi­er to the pub­lic, the pat­tern will go away. The pat­terns will change. And there’s sort of a corol­lary that banks always say is, Well gee, because of this thing we need to make sure every­thing we do is very secret, because once we expose our sta­tis­ti­cal insight it dis­ap­pears. It becomes use­less and the clas­si­fi­er becomes easy to game.” So this is sort of a case they’re mak­ing for opac­i­ty. And it’s not entire­ly false. 

What I’m inter­est­ed in as sort of a math guy is well, is that real­ly true, right? Is Goodhart’s Law a law as in you know, the law of large num­bers. Something you can prove under some rea­son­able for­mal­iza­tion. Or is it not actu­al­ly reli­ably true? Can we some­how come up with clas­si­fies that you can expose to the pub­lic and they’re still hard to game? 

And this is sort of an ongo­ing inves­ti­ga­tion with Brendan Juba and Omer Reingold. We sort of take a kind of a game the­o­ret­ic point of view. We fac­tor in the cost of chang­ing one’s attrib­ut­es in the clas­si­fi­ca­tion process. For exam­ple get­ting, a col­lege degree is much hard­er than buy­ing a bunch of books. So we take the cost of chang­ing one’s attrib­ut­es into account as well as the indi­vid­u­al’s bud­get for any giv­en clas­si­fi­ca­tion task. People are only will­ing to expend so much effort or money. 

And once we have these sort of game the­o­ret­ic ingre­di­ents we can talk about design­ing deci­sion bound­aries or clas­si­fiers in such a way that a typ­i­cal negatively-classified per­son does not have the means or the bud­get to force a pos­i­tive clas­si­fi­ca­tion by any change of attrib­ut­es. And this is true even if the clas­si­fi­er is exposed. So that’s an impor­tant point. So it allows you to inves­ti­gate which clas­si­fiers are hard to game even if you expose them. I think this is a kind of start­ing point to get at some of the issues we’re talk­ing about here from a for­mal perspective. 

So final­ly I want to men­tion a fas­ci­nat­ing paper that unfor­tu­nate­ly I did­n’t coau­thor. But my advis­er did and a bunch of oth­er great peo­ple. So it’s by Sanjeev Arora, Boaz Barak, Markus Brunnermeier, and Rong Ge. And it’s about intractabil­i­ty in the finan­cial deriv­a­tive mar­ket. And the slides are actu­al­ly by Sanjeev. 

So here is a quick recap of the finan­cial crash in 2008. So basi­cal­ly what hap­pened is that the deriv­a­tive mar­ket or it’s total val­ue had become far greater than the world’s GDP. And so the hard les­son that we learned is that even small errors in pric­ing deriv­a­tives lead to huge effect on the econ­o­my just because the mar­ket is so enormous. 

And so what these guys were inter­est­ed in is the sort of intrin­sic com­pu­ta­tion­al com­plex­i­ty of pric­ing finan­cial deriv­a­tives. And here’s an infor­mal state­ment of the fas­ten­ing the­o­rem that they proved. They showed that eval­u­at­ing the fair val­ue of a derivative—a finan­cial deriv­a­tive such as a CDO is an intractable prob­lem. So I don’t want to quiz the audi­ence on what intractable means, but it means accord­ing to Bob that you can­not solve it in poly­no­mi­al time. So even with a sig­nif­i­cant increase in com­pu­ta­tion­al pow­er, even in twen­ty years you can­not solve this problem.

Moreover, what’s even more oppress­ing here is that the design­er of the CDOs can tam­per the prod­uct in such a way that it is com­pu­ta­tion­al­ly hard to detect the tam­per­ing. And these two results rely on the com­pu­ta­tion­al intractabil­i­ty of the Dense Subgraph prob­lem. So this is a beau­ti­ful, well-defined com­bi­na­to­r­i­al opti­miza­tion prob­lem that peo­ple have stud­ied for a long time and believe is quite hard. 

23

And I want to just briefly show you where the Dense Subgraph comes in. So this a lit­tle bit of CS. So this is basi­cal­ly the point of view of the deriv­a­tive design­er. He has a bunch of asset class­es; there are N asset class­es. And he knows that a bunch of them are lemons. And because he’s the design­er of the deriv­a­tive he might have some infor­ma­tion about which asset class­es are lemons. And there’s the buy­er and the reg­u­la­tor who’s like, Well gee, which of these assets are lemons? I don’t know but I want to find out.” 

Now what the design­er of the deriv­a­tive can do, he can cre­ate prod­ucts on these CDOs which are basi­cal­ly just bun­dles of asset class­es. So this is indi­cat­ed by these ver­tices on the bot­tom of these ver­tices on the top. And now basi­cal­ly what the design­er wants to do, he wants to pack as many lemons into as few prod­ucts as pos­si­ble, to get rid of the lemons. And of course this has a pret­ty dev­as­tat­ing out­come on the expect­ed default rate of these tam­pered products. 

And this dense sub­graph, this pack­ing between lemons and prod­ucts, that’s exact­ly what’s hard to find. So if the graph is large enough, then this dense sub­graph that’s con­tained in this bipar­tite graph is actu­al­ly hard to detect. So the reg­u­la­tor will say, Oh no, it’s the Dense Subgraph prob­lem. I can’t detect this.”

So this is just a lit­tle bit of com­bi­na­to­r­i­al opti­miza­tion, how it tells you some­thing about how hard it is to price finan­cial prod­ucts. And not just sort of on an intu­itive lev­el, like fun­da­men­tal­ly, math­e­mat­i­cal­ly there’s noth­ing you can do about it. You can’t fig­ure out what the fair price is. 

So let me wrap up. So here’s my pro­pos­al, again it’s Occupy Algorithms. And what I want to sug­gest is that you guys should help us com­put­er sci­en­tists in find­ing for­mal def­i­n­i­tions and mod­els of the prob­lems that you care about. Because as Bob said yes­ter­day, we’re not going to solve a prob­lem that’s not well defined, that’s not in a math­e­mat­i­cal mod­el. On the oth­er hand find­ing mod­els that’s you know, often most of the work. That’s extreme­ly dif­fi­cult and we need peo­ple who have the domain knowl­edge to to tell us what the right mod­els might be. We can do the for­mal part but we need insight on what a good mod­el should do. So this is an inter­dis­ci­pli­nary thing. 

And if we do that we can make social con­straints a first-order prob­lem in algo­rithm design. And this is actu­al­ly not unre­al­is­tic. This has hap­pened in the past, for exam­ple there’s lots of lit­er­a­ture on privacy-preserving algo­rithm which I find quite fas­ci­nat­ing. And on truth­ful mech­a­nism design, where the goal is to cope with strate­gic behav­ior. So there are areas in CS where yes, social con­straints a first-order con­straints and we can cre­ate more of these cas­es. So with that I want to thank you for your attention.

Further Reference

The Governing Algorithms con­fer­ence site with full sched­ule and down­load­able dis­cus­sion papers.