Samim Winiger: Welcome to Ethical Machines. We are your hosts…

Roelof Pieters: Roelof.

Winiger: And Samim.

Pieters: Ethical Machines is a series of con­ver­sa­tions about humans, machines, and ethics. It aims at start­ing a deep­er, better-informed debate about impli­ca­tions of intel­li­gent sys­tems for soci­ety and individuals.

Winiger: For this episode, we invit­ed David J. Klein to talk to us about machine learn­ing, con­ser­va­tion, and cli­mate change. Let’s dive in.

Thanks for mak­ing the time. Welcome to Ethical Machines. It’s a plea­sure to have you on. Maybe we can start with the obvi­ous ques­tion, could you tell us who you are, your back­ground, and what brings you here, basically.

David J. Klein: Sure. Well, I grew up on a ranch in Florida, and I spent many years sort of mar­veling at nature. At the same time I was a huge sci­ence fic­tion fan, so I’d come back in and take apart all of my motor­ized toys and put them back togeth­er, and watch Doctor Who and read Asimov and Bradbury.

I even­tu­al­ly decid­ed to go in a tech­nol­o­gy direc­tion in my career, although I could’ve eas­i­ly gone in a dif­fer­ent direc­tion. And I went to Georgia Tech. But after a while I became increas­ing­ly unin­spired by the work I was learn­ing about in elec­tri­cal engi­neer­ing. So I start­ed look­ing for a way to keep myself inter­est­ed. So I was tak­ing cours­es in psy­chol­o­gy and genet­ics, and a cou­ple of real­ly impor­tant things hap­pened as I was searching. 

First of all, I hap­pened upon a course called Sensory Ecology. It was taught by a pro­fes­sor at Georgia Tech; the name of that the pro­fes­sor was David Dusenbery. Sensory Ecology is real­ly about infor­ma­tion trans­mis­sion in bio­log­i­cal sys­tems, and how behav­ior and mor­phol­o­gy of organ­isms coe­volves with their infor­ma­tion trans­mis­sion and recep­tion sys­tems. And I was huge­ly inspired by that course.

And so I start­ed look­ing at ways of com­bin­ing Double E with stud­ies of the brain. So the sec­ond impor­tant event was I asked around and I found a young pro­fes­sor at Georgia Tech who had recent­ly come there out of the lab of Carver Mead from Caltech. And Carver Mead and his stu­dents were the pio­neers of this field of neu­ro­mor­phic engi­neer­ing. And so I was able to land an under­grad­u­ate research assist­ant­ship in Steve’s lab, and I was doing research and devel­op­ment on neu­ro­mor­phic vision chips. So we were design­ing vision chips that were mim­ic­k­ing the pro­cess­ing being done in the mam­malian retina.

And real­ly since then, every­thing I’ve done has been in that inter­sec­tion area between neu­ro­science and engi­neer­ing. My grad­u­ate work was in a Double E lab of Shihab Shamma’s in University Maryland, but we were doing exper­i­men­tal neu­ro­science there, study­ing the pro­cess­ing of sound in the audi­to­ry cor­tex. From there on I was a post­doc­tor­al researcher at the Institute for Neuroinformatics in Zurich. So I was work­ing on audi­to­ry AI projects there an audi­to­ry rep­re­sen­ta­tion learning. 

And then I came out to Silicon Valley about a decade ago and I joined a com­pa­ny Audience, where we had the vision of reverse engi­neer­ing the human audi­to­ry sys­tem in order to do a bet­ter job of speech enhance­ment and audi­to­ry source sep­a­ra­tion and robust speech recog­ni­tion. And we devel­oped a chip that went into the iPhone and went into the Samsung Galaxy, and it was a great suc­cess. We were the first com­pa­ny to pull mul­ti­ple micro­phones into a smart­phone. And as these sig­nals are com­ing in they’re first trans­formed by com­pu­ta­tion­al mod­els of the mam­malian cochlea. So not using a Fourier trans­form but actu­al­ly using a fil­ter back inspired by the mam­malian cochlea.

So from there on, I’ve been work­ing on var­i­ous projects in var­i­ous star­tups includ­ing my own. I was using autoen­coders start­ing in 2008 to beat the state of the art stan­dards in video com­pres­sion. And I’ve been work­ing on a bunch of dif­fer­ent projects as a con­sul­tant, adding most­ly deep learning-fueled intel­li­gence to var­i­ous prod­ucts rang­ing from face recog­ni­tion to snor­ing recognition.

So on the con­ser­va­tion side, I was lucky to get con­nect­ed to these researchers at University California Santa Cruz. They were start­ing a com­pa­ny sev­er­al years ago which is called Conservation Metrics. And this com­pa­ny, it was based on their work apply­ing pas­sive acoustic mon­i­tor­ing tech­nol­o­gy to mon­i­tor and help save endan­gered sea birds, most­ly. Over time, I devel­oped tech­nol­o­gy for them and now they have this analy­sis pipeline for all the acoustic data that they get in. So the biol­o­gists who are ana­lysts in the com­pa­ny have the abil­i­ty to build deep learn­ing mod­els to detect endan­gered species of inter­est and get to a more detailed under­stand­ing of how these pop­u­la­tions are doing and how they’re respond­ing to con­ser­va­tion interventions.

And so that’s been excit­ing in that it’s had a very large impact on their work. Their analy­sis through­put has increased by ten times com­pared to meth­ods they were using before using deep learn­ing mod­els. And I think we’re just scratch­ing the sur­face. We’ve expand­ed from audio pro­cess­ing to image pro­cess­ing, most­ly land-based cam­era net­works that are used by con­ser­va­tion­ists today. And there’s a lot of poten­tial. I mean, the vision going beyond that is inte­grat­ing all kinds of sen­sor sources, all the way from envi­ron­men­tal DNA sam­pling all the way up to satellite-based imagery. All of these sen­sor types have a bear­ing on the wildlife con­ser­va­tion and more broad­ly envi­ron­men­tal con­ser­va­tion problem. 

Winiger: So fol­low­ing up from there, the research paper that you pub­lished a while back called ​“Deep Learning for Large Scale Biodiversity Monitoring”. How does this play into the work you just mentioned?

Klein: It real­ly has to do with the vision. So Conservation Metrics, we’ve been solv­ing very spe­cif­ic prob­lems in the con­ser­va­tion sec­tor using deep learn­ing. And it’s great to be able to work with con­ser­va­tion sci­en­tists and their exist­ing projects today and to see what prob­lems they have and how can the process be stream­lined using machine intel­li­gence. And that’s why Conservation Metrics was labeled as a ​“laser” in this recent TechCrunch arti­cle by Shivon Zilis. We’re very much focused on spe­cif­ic prob­lems that exist in projects today.

But there’s this broad­er vision, the idea that we can lever­age these sen­sor net­works that we’re putting out in these remote areas. So we have these things across the world. I mean, we have them in Australia, we have them in Hawaiʻi, we have them in coastal areas of the United States. And there’s at least an order of mag­ni­tude greater need for mon­i­tor­ing. I mean right now we’re using these micro­phones and cam­eras net­works on the ground, but the con­ser­va­tion sec­tor believes that there’s a lot of val­ue that can be gleaned from for exam­ple satel­lite imagery or DNA sam­pling called eDNA. 

If we real­ly want to have a detailed enough under­stand­ing of these ecosys­tems so that we can real­ly engi­neer solu­tions on less than a ten-year run­way— I mean, right now we don’t real­ly have that under­stand­ing. It’s actu­al­ly one of the most impor­tant insights that I’ve got­ten in work­ing with biol­o­gists and ecol­o­gists, is that today it’s actu­al­ly not real­ly known on a sci­en­tif­ic basis how well dif­fer­ent con­ser­va­tion inter­ven­tions will work. And it’s because we just don’t have a lot of data. I mean, these con­ser­va­tion projects—think about try­ing to save pop­u­la­tions of endan­gered sea birds that might feed in islands close to Japan but breed in islands close to Hawaiʻi. I mean it’s a huge area; there’s no way you can send peo­ple out there to get enough data to devel­op a sci­en­tif­ic under­stand­ing of the prob­lems. How these species are being impact­ed by human actions.

So we need tech­nol­o­gy, we need to deploy sen­sors and many dif­fer­ent types of sen­sors to mon­i­tor these pop­u­la­tions and mon­i­tor these ecosys­tems. And we need algo­rithms like deep learning-based algo­rithms that we can use to dis­till insights from this data. Because it’s way too much data. I mean, step one is get­ting the data but it’s way way too much data for peo­ple to look at direct­ly. We have a project in Kauaʻi where we’re detect­ing the sound of an endan­gered bird col­lid­ing with pow­er lines there. And we’ve dis­cov­ered that it’s a much big­ger prob­lem than it was pre­vi­ous­ly thought because we were able to extend the tem­po­ral scale of the mon­i­tor­ing using these micro­phone networks. 

When we get data back from the lab, it’s hun­dreds of thou­sands of hours of audio. It would take a sin­gle per­son ten years just to lis­ten to that, let alone find things of inter­est. So that’s where we’re apply­ing deep learn­ing, where we’re enabling these biol­o­gists through var­i­ous inter­est­ing means to build mod­els and then dis­till these hun­dreds of thou­sands of hours, or many mil­lions of images, down to a small sub­set that they can use in their back­end analy­sis of how pop­u­la­tion den­si­ties are changing.

Pieters: One more ques­tion, like how this relates not just to extinc­tion of glob­al ani­mal pop­u­la­tions but also to things like the state of bio­di­ver­si­ty or cli­mate change? Shat would you gen­er­al­ly say is the impact machine intel­li­gence in a kind of broad sense can have in this area, and is it already hav­ing enough impact?

Klein: The sit­u­a­tion appears to be very dire. I mean, if you look at global.biodiversity[?], the last about half of the world’s ani­mal pop­u­la­tions since 1970, species extinc­tions are orders of mag­ni­tude above the nat­ur­al back­ground rate. And a lot of sci­en­tists are call­ing this the sixth extinc­tion, and there’s no debat­ing that it’s due to human caus­es. And it’s a bunch of dif­fer­ent human caus­es. The num­ber one cause today is not cli­mate change, it’s direct exploita­tion. It’s farm­ing, it’s fish­ing. Then we have pol­lu­tion. Amphibians and birds are being just wiped out. Insect pop­u­la­tions as well.

And you know, glob­al spend­ing on this has increased in recent decades. So the UN has rec­og­nized that this is going down­hill at an alarm­ing rate, and right now we’re spend­ing about $20 bil­lion a year glob­al­ly. But all met­rics are show­ing that it’s not real­ly help­ing. When you ask, ​“Well why isn’t it help­ing? What could we be doing bet­ter?” And there’s not real­ly great answers out there right now. Funding is being dri­ven by emo­tion and log­ic and some mod­els. But it’s not real­ly being based on data. You know, we can argue that cer­tain species that we care about because they’re cute, or we can make impas­sioned arguments. 

And that’s real­ly the kind of thing that’s dri­ving mon­ey flow today. But there’s not a lot of data that’s show­ing us how well we’re doing with a par­tic­u­lar kind of inter­ven­tion ver­sus anoth­er kind. Let’s talk about birds again. You know, should we remove an inva­sive snake or we should we build arti­fi­cial nests? Which one of those is more effec­tive? Usually we just argue for one, we do it, and then many years lat­er we deter­mine if it worked or not. 

So your ques­tion’s about how could tech­nol­o­gy help. So, we do expect that cli­mate change will become the num­ber one prob­lem, pret­ty quick­ly. Because it’s shift­ing habi­tats at a rate that nat­ur­al ecosys­tems can­not keep up with. We’re erod­ing the val­ue of nature. So we can get into how do we mea­sure the val­ue of nature. Actually that’s a real­ly inter­est­ing topic. 

The two things we can do obvi­ous­ly on the tech­nol­o­gy side, one is try­ing to slow cli­mate change. And we can do that by var­i­ous means. We can inno­vate on tech­nol­o­gy for ener­gy; clean tech­nol­o­gy that is much less destruc­tive to our atmos­phere. We can inno­vate on solu­tions for trans­porta­tion that uses less ener­gy. So that’s one side of things, try­ing to slow the degradation.

And the oth­er side of things, where I’ve been more focused, is devel­op­ing sys­tems that enable sci­en­tists to under­stand these sys­tems that we’re mod­i­fy­ing so that as we make con­ser­va­tion inter­ven­tions we can say on a more fine-grained basis how well they’re work­ing. Ultimately, we might need to under­stand these sys­tems so that we can restore them. 

Winiger: I mean, the ques­tion you just raised, how do you val­ue nature. I want to dive into that as you brought it. What is the cost func­tion of valu­ing nature? It seems like a real­ly hard prob­lem to crack. Is there any think­ing around this?

Klein: There’s been quite a bit of work in this area called ecosys­tem ser­vices. It’s econ­o­mists and biol­o­gists, ecol­o­gists, are get­ting togeth­er and per­form­ing an increas­ing­ly detailed account of how nature serves us. You know, what more tan­gi­ble val­ue do we derive from ecosys­tems? And there’s a bunch of dif­fer­ent ways. If you look at how we use bee pop­u­la­tions to pol­li­nate crops and all the val­ue we would get from those crops. The fact that ecosys­tems form nat­ur­al buffers that pro­tect our pop­u­la­tions from storms. The fact that trees take a lot of car­bon out of the atmos­phere and there­fore reg­u­late our plan­e­tary tem­per­a­ture. And many many many oth­er ways. I mean, we even derive pes­ti­cides and med­i­cines from nature. So if you add all that up, we’re cur­rent­ly at an esti­mate of around $125 tril­lion a year of val­ue that we’re extract­ing from nature. That’s rough­ly dou­ble glob­al GDP. 

So that work is going on. But of course there’s a great debate about ecosys­tem ser­vices. You know, can you actu­al­ly quan­ti­fy? Because a lot of peo­ple will say, ​“A future with no nature is not a future I want to be in.” How do you quan­ti­fy life? That’s a real­ly great ques­tion. I’m not aware of work beyond brain­storm­ing. When you start look­ing at the uses of rein­force­ment learn­ing for mon­i­tor­ing and maybe main­tain­ing nat­ur­al sys­tems, it does beg the ques­tion okay, what’s the rein­force­ment sig­nal? What are the objec­tive func­tions here that are being opti­mized? And I don’t have a great answer for that. But it’s a great area of debate and dis­cus­sion because that may be one of the only solu­tions we have.

The approach that I’ve been tak­ing right now is okay, we’re get­ting in these petabytes of data from sen­sors and we’re get­ting that down to a very very small sub­set. But that may not end up work­ing out. The scale of the prob­lem may be too large. I mean, how many hun­dreds of tril­lions of dol­lars are we going to have to spend to restore these sys­tems? I think the much bet­ter approach would be to let these sys­tems care of themselves. 

But what is the objec­tive func­tion? It’s not just the pres­ence of activ­i­ty, like life activ­i­ty. One of the great exam­ples is the con­cept of the eco­log­i­cal trap. So, we have areas like Central Park in New York which were thought to be eco­log­i­cal traps. So it attracts ani­mals, there’s a lot of life there. But there’s not a lot of renew­al of life; it’s kind of a dead end. So if we just designed a rein­force­ment learn­ing sys­tem to say okay, let’s find auto­mat­ic actions that will increase the diver­si­ty and plen­ti­ful­ness of life in a cer­tain loca­tion, that in itself is not enough. We need to have a much more detailed under­stand­ing of what is a healthy ecosys­tem. There’s always a bal­ance there. Today we don’t have a detailed sci­en­tif­ic under­stand­ing. We’re just scratch­ing the sur­face. So that’s why I’m excit­ed about devel­op­ing tech­nol­o­gy that can help us see what’s going on.

Pieters: What would you say, because one of the argu­ments being made, the Singularity will take care of it, Moore’s Law will auto­mat­i­cal­ly solve cli­mate change, ani­mal extinc­tion and relat­ed prob­lems. Or almost like the invis­i­ble hand of the mar­ket will fix it. You know—

Klein: Yeah.

Pieters: So what would you say to these kind of…

Klein: Yeah… I think that’s… I think it’s pret­ty dan­ger­ous think­ing. I mean you know, can any­body point to any kind of tech­nol­o­gy pro­jec­tion of more than thir­ty years that’s end­ed up being accu­rate in any sig­nif­i­cant way, any action­able way? I mean, why do we think this is dif­fer­ent now? Basing our future on a wait-and-see atti­tude, it’s like…it’s just dan­ger­ous. I think prob­lems like this are going to be solved with a lot of work, and work on all these three things: tech­nol­o­gy, sci­ence, and pol­i­tics you know, pol­i­cy. We need to tack­le all these prob­lems in a very method­i­cal and a coor­di­nat­ed way. And the thing is even as we fail—there’ll be a lot of fail­ures, but we’ll be learn­ing a lot, and we’ll be cre­at­ing an under­stand­ing that will be much more broad­ly use­ful for humankind. So the idea that tech­nol­o­gy inno­va­tion is passed down humans from the moun­tain and it’s just going to solve every­thing to me does­n’t ring true.

Winiger: You hint­ed at pol­i­tics and pol­i­cy­mak­ing, and I’m going to frame that as culture.

Klein: Yeah.

Winiger: This notion that large-scale change will hap­pen, it’s tech­no­log­i­cal change and cul­tur­al change go hand in hand. And so do you actu­al­ly see machine learn­ing help change our cul­ture? So our beliefs, our caus­es, our priorities.

Klein: That’s such an inter­est­ing ques­tion. I think that as we get a more detailed under­stand­ing of nat­ur­al ecosys­tems, in part by attack­ing this prob­lem, that we’ll start to be able to have the abil­i­ty to cre­ate these cyborg ecosys­tems. So I would rec­om­mend you look­ing at the work of Brad Cantrell. He’s an archi­tect. And there’s oth­ers like him, they envi­sion this future where we have this kind of con­flu­ence of intel­li­gence in mon­i­tor­ing the envi­ron­ment, and also robot­ics, and also if you look at advances in mate­r­i­al sci­ences, we can start to cre­ate cities that are much more tight­ly inte­grat­ed with nature. Cities where nature more flows through cities and we under­stand how to inter­face with nature in a much more fine-grained way. 

The aes­thet­ic that dri­ves me in that area has been sci­ence fic­tion depic­tions of future Earth. You know, future Earths that are very green, where we have nature inte­grat­ed with cities, down to our ener­gy inno­va­tions are inspired by nature. You know, our archi­tec­ture’s inspired by nature.

There’s anoth­er part to it that’s a lit­tle bit more weird but I think also worth dis­cussing. Because now we have a much more detailed under­stand­ing of genes, so how to inter­pret genes and how to mod­i­fy them. That’s actu­al­ly one of the big impact areas right now of deep learn­ing. And if you look at the work that’s being done in image pro­cess­ing, the gen­er­a­tive and cre­ative art that’s com­ing out of that, there’s a future in which I believe that the inter­face with nature could become a lot more inti­mate at the genet­ic lev­el. So we’ll be able to start envi­sion­ing hybrid struc­tures between humans and the nat­ur­al world that we cre­ate with these gen­er­a­tive models.

Pieters: So this would give a new mean­ing to personalization. 

Klein: Yes!

Pieters: A very dif­fer­ent kind of gen­er­a­tive rec­om­mender systems.

Klein: Yeah. I bring this up as kind of like a vision and aes­thet­ic fuel. I go about my day-to-day being the laser. You know, look­ing at prob­lems and solv­ing spe­cif­ic prob­lems. But as you go along you need some­thing that is dri­ving you in a direction.

Pieters: All these humans are dri­ving a lot of these prob­lems we’ve been dis­cussing. And so I think prob­a­bly a log­i­cal way of approach­ing a solu­tion is social engi­neer­ing, in a sense. I sup­pose using machine learn­ing to influ­ence pop­u­la­tions, maybe we’ll see some of that or we’re already see­ing some of that.

Klein: That’s a very inter­est­ing point. I would love to be able to use tools, and they’re start­ing to mature you know, where we can start to under­stand the whole chain from how ener­gy is derived and how it’s used, and then ulti­mate­ly what you’re using that ener­gy for. Andrej Karpathy had an inter­est­ing tweet a few months ago that got me think­ing. He made some cal­cu­la­tions that showed how much equiv­a­lent wood he was burn­ing in pow­er­ing a GPU to solve a par­tic­u­lar machine learn­ing prob­lem. I think it’d be fas­ci­nat­ing to have a more detailed under­stand­ing of that whole chain. Where the ener­gy’s com­ing from, how the ener­gy’s formed, and how we’re using it.

And so with that visu­al­iza­tion, I think as a soci­ety we’ll start seem more opti­mal ways of liv­ing. The one that we dis­cuss a lot, obvi­ous­ly, is trans­porta­tion. I mean, the fact that cities—well, in the United States in par­tic­u­lar and in China—are just ludi­crous in how much ener­gy we spend get­ting around to buy milk and to work at our desk jobs with no thought about con­se­quences on a day-to-day basis. It’s inter­est­ing to dis­cuss how there’ll be a cul­tur­al change for every­day, smaller-scale mun­dane tasks, once we start to be able to visu­al­ize this kind of ener­gy flow.

Pieters: Maybe shift­ing gears a bit. When you were talk­ing ear­li­er about con­sul­tan­cy for social good, in a sense, or for mean­ing­ful prob­lems to solve. You tend to hear more and more this kind of ​“X for social good.” So what is your expe­ri­ence on run­ning prof­itable social enterprise?

Klein: Yeah, it’s a real­ly inter­est­ing time right now. There’s so many for-good and for-profit com­pa­nies start­ing up. And then so there’s a very healthy debate going on about whether or not that’s a good thing. The rea­son it’s hap­pen­ing is par­tial­ly because folks try­ing to do the right thing, try­ing to imple­ment social good and non­prof­it orga­ni­za­tions, have been frus­trat­ed. I mean, they’ve been real­ly frus­trat­ed by the slow pace of progress and the fact that they don’t have access to the top-quality tal­ent and they spend a lot of their time try­ing to raise mon­ey. And so they see things hap­pen­ing in the tech world that are much more rapid­ly inno­vat­ing through this ethos of com­pe­ti­tion and dis­rup­tion. And I think we have ben­e­fit­ed from that but there’s a lot of prob­lems to solve there still. I mean I think it’s a great thing to try to flesh out. 

The thing is we want some­thing that’s more flex­i­ble. We’re try­ing to find an inter­est­ing way to struc­ture things so that the act of solv­ing the prob­lem gen­er­ates enough prof­it to sup­port a vibrant tech­nol­o­gy inno­va­tion scene, more like a tra­di­tion­al tech start­up. One of the things you have to guard against is mis­sion drift. Time will tell what’s the best strat­e­gy in doing that. I think that it is a dan­ger­ous thing to have mon­ey dri­ving deci­sions now. Because of peo­ple that care the least about the mis­sion and the most about mon­ey will tend to get the most pow­er with­in orga­ni­za­tions unless we can have ways of dili­gent­ly pro­tect­ing against that kind of thing. 

Winiger: Could you envi­sion let’s say the year 2025 where an enti­ty like Google is a major play­er in renew­ables or con­ser­va­tion? Could you envi­sion such a future?

Klein: I could envi­sion such a future, yes. Newsflash: ener­gy is big busi­ness. And right now it’s being dom­i­nat­ed by the fos­sil fuel indus­try. But that’s going to change. It’s going to be some­thing much clean­er, much more effi­cient. And that tran­si­tion will cre­ate a lot of wealth, a whole new glob­al lead­er­ship that has the plan­et’s health much more in their minds I hope I can see in sev­en­ty years or so what that looks like. I think we’re going to be beyond what we cur­rent­ly see in solar pow­er and wind pow­er. We’re going to have much more inter­est­ing cyborg inter­faces with the nat­ur­al world.

Winiger: In Germany they have this amaz­ing trans­for­ma­tion unfold­ing in real time with 30% peak time ener­gy pro­duc­tion now on renew­ables. But the thing that real­ly gives me hope is that half of that renew­able ener­gy is in citizen-controlled hands—

Klein: Good point.

Winiger: —coop­er­a­tive hands. And there’s this kind of silent rev­o­lu­tion of dis­trib­uted pow­er unfold­ing, and dis­trib­uted struc­tures that con­trolled the tech, basi­cal­ly. Which is very uplift­ing, from my perspective.

Klein: That’s some­thing that I’ve just recent­ly become inter­est­ed in and aware of, these kind of dis­trib­uted val­ue chain sys­tems. There’s lot talk about Bitcoin, blockchain. And you know, that’s some­thing that I did­n’t ful­ly under­stand the poten­tial of, these kinds of mar­kets until recent­ly. So I’m real­ly look­ing for­ward to dig­ging more into that and under­stand­ing how they can be lever­aged. You can imag­ine sys­tems where this plan­e­tary net­work of sen­sors is being put into a glob­al, dis­trib­uted CDN. And solv­ing the most crit­i­cal prob­lems with that data will set the price of the data and the val­ue of col­lect­ing data, and the val­ue of solv­ing prob­lems with that data.

I do think that machine intel­li­gence is going to be a big part of this. I think it’s debat­able how much that will add to a human under­stand­ing of these sys­tems. I tend to be on the side of we as humans use machine intel­li­gence ulti­mate­ly to derive insight. These large-scale machine intel­li­gence sys­tems will add an incred­i­ble amount of under­stand­ing of how the nat­ur­al world works. 

Pieters: For instance here in Stockholm, we’re hav­ing a project where we’re actu­al­ly look­ing at wind­mills. Windmills are high­ly inef­fi­cient, in the sense that they pro­duce ener­gy but you don’t real­ly know how much ener­gy it pro­duces where or when—

Klein: Right.

Pieters: —because you need a good weath­er model—

Klein: Yes.

Pieters: —and that’s super dif­fi­cult to do. So a lot of things to win there, by using things like deep learn­ing, mak­ing mod­els much bet­ter, and there­by also mak­ing things much more efficient. 

Klein: Yes. My under­stand­ing is that that’s one of the pri­ma­ry dri­vers right now of the increase… The per­cent­age of total ener­gy in the United States com­ing from wind is being large­ly dri­ven by more accu­rate weath­er pre­dic­tion mod­els. That did­n’t hap­pen by acci­dent. That actu­al­ly hap­pened through pol­i­cy. And that pol­i­cy was multi-dimensional. There was poli­cies to put fund­ing into large-scale com­put­ing sys­tems that facil­i­tate this kind of work. Put mon­ey into fund­ing algo­rith­mic research that can lead to improve­ments in weath­er pre­dic­tions. And the hope was that those would lead to increased uptake of wind pow­er, and that’s hap­pen­ing. So that’s a great exam­ple of this multi-pronged tech­nol­o­gy, sci­ence, and pol­i­tics that can be suc­cess­ful. But there’s a lot more that can be done. We can’t claim any kind of vic­to­ry right now.

Pieters: You men­tioned a few things which kind of give hope for the future. One is cul­tur­al change, that peo­ple will become more aware of these things but then also devel­op more tech­nol­o­gy to address these issues. Better met­rics. More accu­rate inter­ven­tions. And things like restoration.

Klein: Yes. Geoengineering is a scary but poten­tial­ly inevitable outcome.

Pieters: So which oth­er things kind of excite you? Let’s say at NIPS, what are you inter­est­ed in at NIPS?

Klein: I think the area that I am per­son­al­ly the most excit­ed about is actu­al­ly one of the fur­thest away from my domains of expe­ri­ence, which is genet­ics. The tech­nol­o­gy evo­lu­tion in genet­ic tran­scrip­tion tech­nol­o­gy is on a dou­ble expo­nen­tial. You know, so Moore’s Law is this expo­nen­tial rela­tion­ship. Genetic tech­nol­o­gy’s on this dou­ble expo­nen­tial. And it’s now afford­able. Ten years [ago] it was impos­si­ble and now it’s afford­able to ful­ly sequence the human genome and any­thing else we get our hands on. And that’s going to continue. 

So this tech­nol­o­gy is going to be every­where. And deep learn­ing is going to be a big part of that. So there’ve been a cou­ple star­tups recent, includ­ing Deep Genomics and Atomwise that’ve start­ed up to tack­le this prob­lem. Existing play­ers such as Illumina are very excit­ed about the poten­tial. And you know, these star­tups are look­ing at every­thing from drug dis­cov­ery to can­cer diag­no­sis based on very small blood samples. 

We have it being used in pre­ci­sion agri­cul­ture, that we can take envi­ron­men­tal sam­ples like very small air and soil sam­ples and detect dis­ease and prob­lems, and have infor­ma­tion we can use to opti­mize agriculture. 

And of course now there’s not just the analy­sis but now we have the gen­er­a­tive part of that with CRISPR and the relat­ed tech­nolo­gies. We’re just start­ing. I mean, if you look at the tech­nol­o­gy that’s being used today for genet­ic analy­sis with say, deep learn­ing mod­els, they’re much much more sim­plis­tic than what we see in these large-scale image recog­ni­tion sys­tems. They’re bor­row­ing from image pro­cess­ing and speech recog­ni­tion and they’re show­ing that like so many oth­er things, right off the bat we’re see­ing large gains in recog­ni­tion accu­ra­cy for detect­ing how cer­tain drugs bind to dif­fer­ent sites on the sequence. 

You know, it’s been pro­ject­ed that the genomics indus­try is going to increase by ten-fold in the com­ing few years. I think that is true. It’s going to become a huge huge indus­try, and machine learn­ing and spe­cial­ized deep learn­ing archi­tec­tures for genet­ic analy­sis and for genet­ic edit­ing are going to become a thing in the next few years.

Winiger: I mean, the cost decrease is obvi­ous­ly the most promi­nent sign of this unfold­ing rev­o­lu­tion, real­ly. But as well it opens up some broad­er, ter­ri­fy­ing sce­nar­ios where you can start the gene dri­ve from the com­fort of your bed­room and prob­a­bly try to mea­sure the impact on a large-scale biosys­tem on your deep learn­ing mod­el at home rather than some­where on AWS. But you only can get that far. These mod­els prob­a­bly will, on the dark side, become a real­i­ty as well, then.

Klein: Yes, yeah. There’s a dark side to all of these things. These pow­er­ful tech­nolo­gies, they have such destruc­tive pow­er if used in the wrong way, inten­tion­al­ly or unin­ten­tion­al­ly. And so how do we address that? I mean, we address it by under­stand­ing, by a hands-on approach, by dis­cus­sion and decid­ing togeth­er as species where we should be apply­ing our ener­gies and how we should be using things. And to have as much trans­paren­cy as pos­si­ble across the board.

Winiger: So are you a real­ist or are you an opti­mist, or… What would you call yourself.

Klein: I’m def­i­nite­ly an opti­mist. Yeah. I am an opti­mist. I’ve changed over time. When I was younger, when I would go into kind of a med­i­ta­tive state I would envi­sion things com­plete­ly falling apart, and poten­tial­ly quick­ly. But as I’ve advanced in my career and I have the abil­i­ty now to talk to pol­i­cy­mak­ers and talk to peo­ple in tech­nol­o­gy talk to peo­ple on the ground doing the work, I’m a lot more opti­mistic. I see that at least in the peo­ple that I have come in con­tact with, and admit­ted­ly that’s com­pared to glob­al pow­er struc­tures it’s a very small slice. But I’ve become opti­mistic. I can see that devel­op­ers in the future will how so much pow­er to imple­ment change. You know, we tend to be a lot that strives for sci­en­tif­ic truth and opti­miza­tion. And I think we will col­lec­tive­ly decide on opti­miza­tion for good. And the thing is you know, good is not an objec­tive thing. It’s some­thing that we all have to con­tin­u­al­ly revis­it as a species.

Winiger: If you made it this far, thanks for listening.

Pieters: And also we would real­ly love to hear your com­ments and any kind of feed­back. So drop us a line at info@​ethicalmachines.​com.

Winiger: See you next time.

Pieters: Adios.