data.bythebay.io: Hunter Whitney, UX and the Changing Relationship Between Mind and Machine
um thanks for the opportunity to present I really enjoy Jeremy's uh talk was very inspiring and I think it um reinforced to me some of the things that I want to talk about today from my perspective and the perspective of I would say ux are there ux designers in the audience at all okay there's a good there's a brave soul here from ux um so um my talk is the coevolution um coevolution the U ux and the changing relationship of human intelligence and um and machine machines AI let's say um there's a bit of a background but just a will add a little bit to that um so I'm ux designer um my main interest is health and medical uh data intensive uh work uh I've gotten more recently involved in cyber security which has some very interesting and somewhat similar problem sets and uh when I was invited to talk today I was thinking about um what I do ux design which I would say is a one way of putting it is it's helping to uh strengthen and improve the relationship and understand the relationship between humans and the machines they work with the computers they work with and as Jeremy was pointing out that is changing at an amazing um clip right now and in ways that are sort of mind-boggling for me as a ux designer I think the most important uh by far significant uh development is this what I'll call artificial intelligence as a umbrella term but uh deep learning all these sorts of Technologies and approaches that are emerging you know thinking about those in relation to um human beings and how we're going to create interfaces or interactions that are going to really work in the future so for this this talk I was reading an article about um humans in a loop uh which is a term referring to thinking about improving machine learning with humans sort of getting that extra The Last Mile the last little bit really kind of fine-tuning that system and it's a super interesting that's anyway as as I understand the term but it got me thinking more about uh other Loops the sort of larger Loops the human uh intelligence and machine intelligence loops and what are those going to look like uh and it seems like to me that those Loops are going to get much more um tight intertwined and subtle it's going to be harder to differentiate I think right now people have a lot easier time knowing what the machine is doing and what I'm doing but even this example that Jeremy mentioned this morning about um suggesting uh email responses and all that you know who is doing that it's the machine but it's it could have come from you it's getting blurrier and blurrier all the time um so from my perspective uh and I have a little bit of bias I suppose but the the crucibles of this sort of evolution are the interfaces and interactions um of which there are many and that these interactions are going to drive um to a certain extent the development of this relationship between human intelligence and Ai and um it really is becoming more and more of a two-way street that people are looking looking at displays um interacting with them and responding to them by the same token uh the machines that they're looking at are looking back at them in a sense they're studying the users and figuring out what they need what they want and to in their own way basically uh adapting to them so it's sort of like a you know we're we're looking at each other and in much more kind of equal footing and the each of those interactions drives probably further development in one way or the other which we'll get into in a minute um so I was reading an article last night and there was a a phrase that really struck and I think is really true organisms or algorithms um you know and I you think about it you know we're basically we're set a code sets um that are uh primed to help help us you know adapt to an environment there's a certain level of so this is uh we could all probably could tell uh double helix DNA um so that's got the the codes that tell us um you know tell us what to do basically or how to operate and how to respond um there's a certain level of stability in that code um and predictability but there's also room for variation so that we can over time adapt to changing environments and I think the same thing is true obviously for for um algorithms and computer code that there are some there's a basic structure there but over time it evolves and adapts and more and more it's adapting to um the environment it's adapting to things that we don't really directly manually control so I just think there's an interesting parallelism between these two um the biological organisms have been adapting primarily through environmental pressures and selection pressures uh whereas um uh in the past I think more so uh code has been really much more directed now it's also adapting to different environmental pressures and behaviors of the people who are using it so we've got these two let's say entities or organisms of A Sort um that are now in this Tighter and Tighter relationship and I have a biology background and so I was thinking about um this idea of co-evolution and it and it sort of un Unleashed a lot of thoughts in my my mind about this whole subject so um basically co-evolution in in simplified form is when two or more species interact with each other and reciprocally change uh each other so in this case um this is a I think a buff tailed uh sickle Bill hummingbird um and you can see it's got a very unique shaped bill or beak um and then there's this flower of which I not sure the pronunciation and I don't recall the exactly but uh that's its partner basically in in evolution and you can see that there's quite a um almost as if it were designed um I don't believe in intelligent design but uh you can see that the shape of the beak and the shape of the flower are very um suited for each other and the idea is that that the the bird is able to get the nectar from this flower very preferentially it's very well suited to get it and that by doing so the flower is able to PO spread pollen on this bird which can then pollinate other of its uh plant bre so in coevolution uh there are a number of different types of relationships they don't always have to be happy good relationships they can be um they can be bad for for one of the partners in that relationship um so this is a kind of a basic uh kind of grid of of how that works so you have um uh one organism benefiting and the other one not um maybe like a host and parasite sort of situation um that's one thing that's antagonism um you've got uh competition say different Predators or Predator prey and um mutualism uh which is both organisms benefit from the uh interactions and the and the the evolution that happens between the two of them um um as one example of that uh there is a a little fish called a a GOI and a and a shrimp um that's pretty much blind they're completely different species very orthagonal to each other in most respects but they actually sometimes will take up residents together inside a a little a little space um and the uh the the GOI has site and it will allow the um shrimp to go about and do its thing but it will when the shrimp is in trouble say if there's something going on in the environment the GOI will see that and get the shrimp protected the shrimp in turn helps the um the GOI by providing closure and bringing in food so there's sort of a reciprocal positive relationship and I think for ux design um those kinds of interactions are the things we want to have happen we want to have these uh these positive mutualistic react uh uh sort of uh encounters and Evolution at least drive towards that and less potentially of the sort of uh uh competitive uh uh computers winning and humans losing kind of scenario I think mutualism is very possible in a lot of cases so as I said when I thought about uh coevolution a lot of ideas kind of Spilled Out and um I wanted to talk about a few of them um I'm still thinking this through um but I think it provides a framework for a lot of different directions uh for thinking about the relationship between uh humans and computers um I'm going to talk a little B about feedback loops uh different types and uh what that might mean stimulus response um you know how people react to the output of machines um relating to that sort of learn behavior um and how that can work work for us or against us um the Red Queen effect which is uh basically something which I'm noticing a lot in the cyers security space um and it's a problem of this uh rapid Evolution essentially um mutualism overdependence and over specialization and then thinking a little bit about machines and biological evolution so feedback loops um there are many typ one classic example is a predator prey relationship and one thing that's really important about this sort of very basic relationship there needs to be a certain type of balance um if it's if the advantage goes too much to say the Predator um they'll wipe out all the prey and then ultimately that will wipe out the Predators so while the relationship may change over time and the Predator may be in the ascendant at certain periods or the prey may be in the Ascend in other periods there has to be at least enough to keep the system going so uh that's there's a feedback loop involved there that um can either be in balance or get out of control this is from um an example of a positive feedback loop and uh the dangers potentially of a positive feedback loop this is an example um uh from a book that I really like called Universal principles of design and it was talking about um feedback loops and unintended consequences so they used an example of football helmets and in I believe the 50s the helmet was very soft and there was a lot of head injuries so they said let's come up with a better solution for that so they came up with the plastic helmets that we all are familiar with um which makes sense it's a better technical solution ution um but then there's the human behavior aspect of that which wasn't necessarily factored in and that is that the players felt more protected so they were therefore um more inclined to take greater risks greater risks meant that the injury rates actually in some ways went up and maybe the severity of the of those injuries went up so it's a kind of a a feedback loop that we don't necessarily want and then so maybe we make the helmets harder and all this and then they people take more risks in this particular case um it may be that it's not just continually improving the currently existing design of the helmet it's thinking about a different helmet altogether or a different approach and actually with the news about concussions and so forth in NFL that um uh that that's going to that's going to change there's going to have to be a rethink that there's a limit to this cycle of just improving the helmet we can sort of see that thing there's a time where maybe it's just going to drop off we just can't we're not going to accept um traumatic brain injury and so forth uh that we have to find a solution technically or behaviorally uh also negative feedback loops um you know thinking about how to time things so that the human interacts in a way that it makes sense this is a little bit old example I think at this point point but uh uh you know using Segways and sort of adjusting Segways so that um so that people are able to control them better and not overcompensate underc compensate same thing is true for say fly by wire um Aviation that the machine helps to sort of Tamp down overreactions uh to the technology by um by humans um I think yet another kind of interesting example that's related to that is the idea of antiock brakes um it's a great technical solution but there's um there's a habit especially of the of the maybe older Among Us or um some drivers when you um anti-lock brakes essentially um you know the machine is handling some of the braking um and so when you apply the pressure to the brake pedal you should keep a consistent pressure and let the machine help out I think some of us were conditioned back in the day to pump The Brak pumping The Brak is not the right solution but that's a sort of an old Behavior so in a sense what we're doing is we're conflicting you know our relationship our expectations what we predict the system's doing is not what the system is doing so we have to sort of relearn and and in another way the system has to help us maybe to to do the right thing in spite of ourselves so that it's a maximized system so um feedback loops in data and uh intensive applications are happening so feedback loops between machine learning and human intelligence human analysts this is an just a pretty basic example of a say looking at a a fraud and a network graph that's been generated by some data data that's being presented in this visualization and on the right the um human analysts can then say yeah that makes sense or or it doesn't make sense like this is um a very quick and dirty way for the human to be in the loop and and put in the human perspective judgment on what the what the machine is seeing and then the machine takes that information and then uh you know depending on what the the human has specified there may reinforce this or may remove it say this is not a a connection that matters or it's fine I know those people so that's an example of a feedback loop now um I look at this kind of thing as as useful important and necessary for the time being although I also think it's very transitory I think there's going to be less and less of this kind of of thing it's still quite manual looking at this sort of visualization and then putting in your comment about that and then doing another one and another one um I think the machines are going to be much more Adept at figuring this stuff out and just coming up with the right answer immediately but in the time this is a a necessary step so in feedback loops um you know we think about the Predator prey or the human machine as you know the the pretty much those are the the the important characters but it's also the environment to which they operate so in the Predator prey relationship in the natural world um there could be some sort of alternate um uh force that upsets the balance of relationship between those two so maybe the um the coyotes are all shot by by hunters um therefore that that's a perturbation in that feedback loop and the the prey animals uh explode in population and you have then potentially a crash so it's not just considering the feedback loop uh between the two kind of main actors but the environment in which they operate in so so for example in this cyber security type application I was looking at we're just looking at a minute ago um the feedback between the human computer is one thing but also thinking about more broadly where's the data coming from uh you know what is a person doing with this data what action do they take it's not just about that immediate interaction of this is good or that's bad to the computer it's what do I do with this information so it's an environmental perspective as well as um um a perspective on just the the two actors so because there's a increasingly inwin intertwined relationship interaction between humans and M and machines we're just looking at there's this machine collaborating with the humans saying here's a a network visualization of potential fraud and the human is saying yeah that looks good or that doesn't look good this is um happening all the time on a um kind of very intense basis and I saw this quote and I I kind of I think about this um this is a sort of a subtle co-evolution that we're spending so much time kind of in this machine world or some of it say security analysts or medical researchers that um potentially we start to think lose the way we think about things or solve problems so a security analyst may have been doing something a certain way they don't necessarily know how they do it but they do a certain way and they get the results they want but now and more they're sort of looping into this sort of kind of um programmatic way of doing things partly as a result of way ux designers and others have designed these systems but a lot of times the people who designed the systems that these analysts are working on are data scientists ux designers and others who don't really have the domain expertise um which is a something that Jeremy brought up and I want to get back to that um they don't have the expertise and they don't have that way of thinking about it so basically they're in a box in which that's been created by people who are not them or the way or their sort of way of thinking about things so I think there is a danger that we become more kind of uh regimented in a sense in the way that we think it's not an insolvable problem but it's just something to to consider um I did a I started doing a little grad work in Neuroscience I love it I think it's really interesting but at least in my belief um we don't really know that much about the brain yet we're learning more and more all the time but it's still it's quite a black box there's still much to learn we're learning about computers but um but again I think with this all uh deep learning and Technologies like that we don't understand necessarily what's going on under the hood either and now we've got the comp combination or Collision or um interaction between these two things which we really don't understand that well at all um you know what is the effect that uh the machine display the machine is having on our cognition our perception our thinking um there's let's just say there's many studies who knows what it all how it all turn out but almost certainly it's affecting the way we we think in certain respects or how we perceive things it's changing us and we're changing it um so say one example of that is the sort of idea of um we get something we put something back out uh right now I work in areas where there's a lot of data flooding people there's no way that they can keep up with the the volumes basically um so they look at dashboards data dashboards that have a lot of information most of which is pretty much in um not actionable in itself they're just love to sort of figure that out and they've also got a million other distractions along with that and um I think one of the things as far as like making evolving our relationship better is dealing with this thing called alarm fatigue or the alarm problem which is basically overloading um humans with too much information to process and do anything meaningful at once machines are great at going through the data and pulling up a lot of interesting anomalies and interesting patterns in the data but figuring out which ones of those to present in the in the right time is a problem at least currently um you know probably many are familiar with this but if you ever been to a hospital and you you go to an ICU and you hear all kinds of buzzers and alarms and all kinds of noise going on and um I think human beings by their nature they're they're evolved to tune the non-critical things out essentially they tune into the things that are really important but when there's a lot of things that are all saying they're important you end up tuning all of it out so in the case of um you know hospitals and medical devices U monitoring devices you know the it turned out in some places for example that the more alarms that were going on the worse the response time time was probably not a surprise um and you know basically you shut off the things that really matter so if there's an alarm that actually counts um you don't hear anymore because your brain is basically habituated to it which it as it should um we could we can't respond to all stimuli all the time that would that's not a very adaptive um uh not very adaptive in the natural world we need to focus on what's important there's a lion over in the bushes as soon as we see that know know it and and and respond the best doesn't matter what else is going off there's a bird tweeting but in this in this kind of new world there's a lot of things crowding for attention um so I think about design and start with a simple example um smoke detectors we uh we kind of know what to expect we know what we want from them if there is smoke in the in the air we want them to make us aware of that doesn't always work sometimes you burn your toast and it goes off but basically that's what we expect and it was designed that way um so smoke loud noise maybe a red light or whatever it's that's a signal that's that's important information we could have designed the system the smoke detecting system to just say everything is fine every five minutes we could have said that's cool that's good you know um uh that obviously would not make a lot of sense it wouldn't work too well um over time because we'd habituate wouldn't even hear it anymore and it and for other reasons we don't there's nothing we should do about it um now we've got a lot of systems that are detecting all kinds of interesting data out there and sort of showing it to us a proliferation of detectors and uh the the problem is that a lot of them at least in this in the work that I do I see dashboards that show this sort of like everything is fine everything is fine everything is fine or not important information on say this dashboard so um I end up um trying to sort of remove that and reprioritize so you know I dashboards that look kind of like this or data sort of like this where there are some very useful information in there but there's also a lot of things that are really unimportant um so it's thinking about at the interface level how to make that information prioritized and accessible so you show everything but you show in a way that actually makes sense from a user human perspective um and uh as simple and obvious as it seems to be if you know I don't think it's practiced as much as it should be um so I was talking about um learn behavior and things like habituation and learn behavior and what to expect like from a smoke detector you hear a loud noise you you generally are going to think there's a problem I got to get out of the building or at least you're going to look into it it's a learn behavior um which is great that's kind of what we want but in the again in in a lot of the work that I see um learn Behavior can be kind of a double-edged sword um especially based on visual cues so this is an example of um co-evolution uh uh and mimicry basically the there are certain kinds of butterflies which are poisonous or toxic to birds uh they don't hide the fact that they exist they don't hide the they actually want to show show themselves with bright colors they want to say I'm here if you want it go for it but you're going to get very sick or die if you do um so that's a very powerful adaptive technique um other butterflies have quote figured out that if I look like a poisonous butterfly but I'm not a poisonous butterfly I can get a free ride here I can I can avoid predation and I'm relying on the bird's kind of predisposition to avoid doing that to eating the poisonous butterfly so they're Free Riders uh the problem of course not for that not for the non-poisonous butterflies but for the birds is if they eat the the non-poisonous butterflies they learn that their that color doesn't really make any difference the next time around they see another one that looks like that they'll eat it and get sick or die and so will the butterflies so it's a bad outcome for for at least two creatures there um and sometimes I feel like in a strange way way the same thing is happening with uh data representations data visualization or showing data people kind of learn to trust data and they see charts and graphs they see dashboards and they kind of believe them without much question without much challenge they're just there they look slick they look good they look like there's they're quantitative um and I've seen this more and more you misused um so you may have some valid data say on the left for something and then you've got something that looks exactly like it on the right which is really poor and you consume the thing on the right you're going to be a sick bird basically um basically you've you've been habituated to trust these things because they look the same but they're not the same so um it's a great evolutionary tactic on the uh part of people who may not want to do the the full work or have kind of bad purposes for their information I guess fishing would be another example as well we're habituated to opening things looking uh looking at someone send us a link and clicking on that so that's just a sort stimulus response and a lot of Bad actors are out there kind of using that sort of immediate learned response so how do we help people maintain a certain level of convenience and um and flow in the world without uh having them eat the bad butterflies so um another uh kind of aspect of coevolution and a term that comes up in that context and is also true I think in in some of the work I'm doing is this the idea of the Red Queen effect which is from Alice and Wonderland and basically the Red Queen tells Alice it's all kind of looking glass world and so everything is kind of backwards in Reverse so um so the Red Queen tells Alice that they're running running running running but they're not getting anywhere and um basically she say like the Red Queen tells Al you have to run as fast as you can just to stay in place and um I feel like that's true for a lot of things especially in the world that um I'm inhabiting right now which is cyber security we're running running running to try and keep up with um say the Predators out there in the world and um we're we're not even keeping up we're actually falling behind um so um you know we're using techniques like machine learning uh to help kind of look at all the network traffic all the anomalies Network traffic that might be suggesting um potential attacks but there's no way to keep up there's absolutely no way to keep up with that so um kind of look towards um machine learning uh with with hope that maybe we can get beyond that that but of course the Predators and the prey in the situation both have access to technology so uh we'll see but it seems like kind of a dead end in a sense or sort of a static with the um with uh the attackers basically always having the advantage um so you know I kind of look at uh the design challenge right now is thinking about this sort of very integrated embedded system where you have in the machines you have deeply embed you try to embed some level of human judgment and you know it's already in there because people created the system at least initially and the flip side you have um you know human beings and looking at the thing through means like data visualizations um and interpretation and then feeding that back to the to the machines um the the thing is that this is not static I think every design that I come up with now probably will be Irrelevant in a couple years and I actually I hope it is um I think there are fun problems to solve but the at the rate at which design um the technology is changing um the designs we working on right now I would expect to be Irrelevant in the say relatively near future so you know there may be some sort of common kind of goals but I think the means to get them are are not going to be the same and so you know um designing for systems that really amplify humans is not going to be a single interface design it's going to be having the machines basically look at us and understand and adapt to us to a certain extent to observe how we're doing where we're making our mistakes or maybe we're doing things and kind of raise those those points so um good it's still working um basically it's more I almost think of like a human collaboration sort of style where um the machines are actively paying attention to us and saying uh you know what are you doing here or yeah that's a great idea um you know that that the machines are kind of good listeners you know that they they actually pay attention they're not just some sort of like dead thing there that we enter information into um that they challenge us that they present information that we mean not have thought of so we're not just sort of looping in our own um kind of stuck way of thinking about things uh and push things out offer suggestions or predictions or kind of other ways of looking at it so I definitely U I'm more and more thinking about not human computer interaction but sort of like a quasi human quasi human interactions if I could put it that way and that the interactions will be guided more and more by the experience in fact I think user experience design has never been a term I've Loved but it's a term of art right now but I I think about it more as sort of user computer experience design because the experience the computer has of the human is going to be important as as well as the other way around so if the human doesn't the human perceives something on the screen and or misperceives it puts in bad data that that's going to be bad experience for the computer bad experience for the computer is going to be a bad feedback back for the human so it's a a vicious cycle in that sense or potentially could be a virtuous cycle depending on the quality experience the quality experience can be guided by good design by making sure that you know it's as low low um little manual input as possible low uh potential for um errors as possible and um and and maximizing things like challenging assumptions or making us think about things in a different way um I talked about cyber security as being um basically it's it's a it feels a little bit Bleak right now um the the the the balance of power I think is in the hands of attackers versus Defenders um but you know in in biology and evolution uh asymmetries and sort of problems create new opportunities to for new types of species in cyber security um there are um a proliferation and abundance of tools now to help defend um probably most of those will go away or may not survive but there's all different ways of approaching the problem by um offering these different sort of prototypes out there just in natural world different animals will uh be born some Will Survive some won't uh this proliferation of different types of of data intensive applications will um lead to some interesting filling of niches and so forth um I don't have time to go through this slide too much but um this is a uh something I saw a couple of days ago I thought was quite interesting in relation to sort of um um diversification of um different types of human computer interactions with Mach learning and basically it's showing how in this um in this uh Matrix here this grid um that certain types of uh areas will be um uh will have different uh importance like so for example you want your machine you want your your reaction interaction with the computer to be predictable and also you want to make sure that um the costs of the mistakes you know are within a range that's comfortable so if something is unpredictable and there's a high risk of mistakes then probably at least for the near- term you're not going to want to have 100% the machines doing it if it's low cost um pretty predictable that's great for machines over time that Dynamics Dynamic is going to change um yeah I want to take time for questions so I'm going to skip to the next uh I do think or philosophically anyway that there there is a potential of of dependency and over specialization that we can function in the world without the machines that were creating in a way they augment us make us super powerful but also um there's a certain fragility potentially in a system like that so if this specialized U flower the mo and moth Dynamic where the moth has a long um tongue to feed if that's gone maybe this flower is going to suffer as a result or go extinct or vice versa so there's always a danger and specialization and dependency um maybe the last thing I well the last thing I'll say on that is that uh that there's actually you know I talked sort of analogically or metaphorically about coevolution of uh the relationship between humans and machines but I think there's also sort of a very direct kind of impact that machines will have on Evolution potentially which is um now with things like Gene editing and uh you know Nanobots that go inside of our our our body that can go inside of our bodies and clean our Our arteries and so forth things like mate selection match Okay Cupid that that there's going to be actually kind of a direct relationship between biological evolution and and the ways that um we develop our Computing systems I have no idea but I just feel like um that along the way that humans um will have a say in how the outcomes of these are I don't I think right now we don't understand the systems that well we don't understand the brain that well but we can at least interact with these two systems we can manage these systems a little bit through interactions and through interfaces so as machines become better equipped to take care of themselves I think maybe we can spend a more time than we currently do and thinking about those relationships and managing them what do they look like um the outcomes I don't think are deterministic I don't I don't know this dire predictions of the machines taking over or losing our jobs I just think we can shape the direction and ideally the main outcome of all this will be that we have a a very powerful um liberating experience uh that helps cure cancer that helps Elevate art all these things are possible but I think the way to get there is by thinking about the interactions over time and what we do over the next five or 10 years in terms of interface design thank you if there are any questions [Applause] yeah um um I'm actually your antiock bre system uh is kind of a good example uh for uh what kind of system we should design I think instead of uh react to um the situation you slide or you the machine we have pedal for you actually should be more collaborative as you said depend on the person who is doing maybe push down all the time or maybe he's just jerking the car already the system if it's smart it should react and I think then become more collab collaborative in a sense right I think probably that's the way we can think even for human with a machine in the future when machine is capable doing a lot of things maybe instead of replacing people make them assist the people with different skill to be able to perform uh the same task of the same quality that we need then maybe for the sake of we are doing something machine is capable of doing at least the s the word is in good order not going to have a disaster because then a lot of people lost job then you will have a disordered uh world that's my thought yeah thank you very good formulation that's sort of the case that I'm I'm I'm trying to make here at a high level else okay well thanks very much appreciate it f