 All right. Thank you everyone for coming. I think we're ready to start the conversation My name is Dalmo I work for a company called work day and I do machine learning applied to financials over there The opinions here are my own there's not the company's opinions But we're gonna have a lot of fun looking at what is to come. Let's Start by trying to build a little bit of trust if I say Can you give me what's in your wallet? You don't have to answer that Yeah, you don't have to answer that it It is not a fair question Because I'm not giving anything in return and I'm asking to get what is in your wallet You don't even know me and you don't have any any trust I haven't even given the price for what I'm offering and I'm offering you no value yet. They're just asking something So it's not a good way to start and build trust Trust without context say, oh you trust everyone. It's like sometimes it could happen, but it is rare It doesn't come by very often Right For example For example if I go in and do something a machine learning model says This transaction has been flagged And I say Why is it flagged? Well, because the machine learning model said so That's not a good explanation, right? We can see over here We have to get more context. You have to be able to explain why was it flagged? Can you tell me what fields what aspect of the transaction were flagged and why? Let's take a look at example imagine that we have a record a transaction physical transaction record and There are three fields that we're going to classify over here and This one it gave a confidence score of zero point one Let's say it ranges from zero to one one being a hundred percent zero being very low confidence And it came back with three classifications one two and three How can we try to figure out why The result was so low why the confidence score was so low We can do a trick Not really a trick but one technique It's a we remove the first field and we run the inference again Now we see the result was a little bit better from zero point one points to zero point fifteen Well, we could say Maybe the first field is the one causing the problem Now we see another one. We remove the second field instead of the first one. We put the first field back And we've removed the second field Now we see the school the confidence score went went to zero point nine two which is much better, right The last one we've removed the third field we have to the operation has to be complete We have to see the whole transaction taking place testing with each one of the fields and see what is the confidence score That comes in the back right And we can see over here that this row seems to provide the best result and The explanation over here we can We can see that the field of the classification number two was the only one causing the problem and we can say Well, the confidence score was a bad one Because the classification two was wrong That is a better explanation. That's a good explanation to why the confidence score was was low and One way for you to start building trust That was a good example if we had said the The first the second row over here the one where we removed the the first classification and if we had said That is our best answer That's not good enough because the confidence score is to zero point fifteen and We would not build trust that way even though we had given an explanation But it would not have been a good explanation So there would be no trust to build over there We're gonna go back to top the topic of trust in a moment, but first we're gonna take a look at game theory We it's gonna be a key component in what we're gonna be discussing today, so game theory is a mathematical framework where To model the in analyze a situation where two people or enter two or more people are interacting with each other In many cases the outcome of the next move depends on the previous move Chess is a good example for that Depending on your strategy when you make a move then your opponent Even though may have a strategy, but may change that strategy as it goes because of the move you made Let's see another example that is Very famous it's called the prisoners dilemma if you guys are not familiar with the prisoners dilemma Let's take a look at how it works Let's say there are two prisoners but you don't have enough evidence to convict either of them and Each one of them you're speaking with each one of them in a separate room They are not seeing what is happening to each other. They don't know what each other is gonna be are gonna be answering They have each one of them has two options and I'm symbolizing each one of them with the color orange or The color blue or green how it's the rendering over there The first option is None of them talks And if none of them talks, let's say that each one gets a sentence of two years in jail Right, they say okay. I'm not gonna talk each one gets a sentence of two years in jail but one of them Can come back and and say I actually am gonna talk because a deal was offered to me that if I talk Then the other I I go free. I am set free and the other person Is gonna spend 12 years in in jail Now we have an incentive over there. Now the person is thinking Should I stay quiet? What if the other person talks and I don't talk then I go to jail But the other person doesn't So my incentive is Do I really trust the other person? Or should I just Talk and And whatever the other person does well the other person does Then you think Okay, I have the the two incentives over there the the The person on the other side is gonna be also thinking oh should I talk or not talk Eventually it comes to a thing called the Nash equilibrium Named after John Nash and if you guys have seen the movie a beautiful mind it It tells like a great story. It's a highly recommend watching But if both of them talk Each one gets seven years in jail Right, so you're gonna risk Do I trust the other person completely? Then each one spends two years in jail Do I stay quiet and then The other person talk and I go to jail the other person doesn't or do I speak in the end the incentive is Everyone's gonna talk because they'll try to minimize the time spent in jail Not talking comes with a potential Penalty of spending 12 years in jail versus if I talk I spend only seven years so A lot of game theory sometimes is on non-cooperative games This one is non-cooperative because each one is compete if one of the prisoners Is competing for their own freedom or minimizing the time they spent in jail There are other modalities on the cooperative games Let's talk a little bit more about that so Let's say there's an actor And you have on one side Something not trustworthy and on the other side The other end of the spectrum you have a situation there You considered to be trustworthy The first question that comes to mind is What is the threshold for being trustworthy? Right we can put a Bar over here and say is this a good threshold? I don't know Maybe We can try to move this a little bit over here to the bottom and say Let me be too low, right? It's it's a very low bar to to gain trust of of someone You can be very Uh Sorry, I'm gonna go back over here That one on the other hand is Very high bar you say oh, I need Something very a guarantee of some something Before I can see the trustworthy so But perhaps we can compute What is that trust? um For being trustworthy How do we go about doing that? Let's build a Trust game So there are going to be two actors over here. The first one is a thruster someone Who is trying to gain the trust? The second one is trustee. So the trust is trying to gain the trust from the trustee But there's something interesting over here. The thruster Has to be trustworthy If you try to act as someone who is not trustworthy You you already started this game in the wrong way, right? Um, the thruster also also has to be trusting It has to be willing to trust The other person But the same is not true for the trustee So the trustee does not Have to be trustworthy For example You can have do business with someone Who supports a different sports team from yours Right You know that their sports team is not as good as yours But you still can do business with them So the trustee doesn't does not need to be trustworthy But what you're trying to change Is for them to be Not trusting To become trusting you want the trustee to trust you And that's how you want to start building trust I don't know if you guys have seen an american show called whose line is it anyways It's a very funny show and they say that The points don't matter And that's where we're going to start over here with we're going to start with a million points It doesn't matter you can be a billion points or just one Whichever number you want to pick Well 42 42 is a good number too And we're going to start playing a game so This Game over here, we still have our two actors. We have the thruster And the trustee and you see that Underneath each one of them There's a pie and we're going to Either increase the size of the pie Or reduce the size of the pie And that's how we're going to start Building trust So the first one the thruster Has to create something of value And i'm symbolizing with that letter v on top of thruster Over there Has to create something of value and be willing to offer that value to the trustee How is it going to do that? So it's going to give that value and we're going to do a little bit of math over here But multiplication is the most complicated part of it. So Bear with me So the thruster is going to do a remittance Just that's a just a funny word for sending Value to the trustee But you notice a little a smaller p over there is Not all the value that was generated by the the thruster Perhaps not all the value can be sent to the trustee Imagine the value being a product many of us working software companies and we build a product and The offering sometimes has tiers that you have to have a free tier or a page tier enterprise one Or maybe the trustee doesn't use all the features of the product So there's usually just a portion of the value that is sent To the to the trustee or maybe it's consuming 100 percent So that small letter p over there is going to vary between zero It means that you're going to be sending no value or all the way to one Where you're going to be sending all the value that you produced So and you're going to send that value To the trustee The moment that you send that you saw the the size of the pie Underneath the thruster got a little bit smaller because it's giving some value away But of course the thruster doesn't want to see its pie shrinking to to nothing Let's deal with these dynamics and see what's going to happen over here A trustee on the other hand has to receive value Perhaps in a greater magnitude than the value that the thruster is sending to the trustee So let's see symbolize this letter k on top of the trustee as a magnification value It's saying how can you receive more value than the thruster Sent to you. Let's see an example Imagine a book An author writes a book. Let's say a great book When you read it You become richer by reading that book. So for you the value that you got from reading that book Was bigger than the value that the thruster sent to you the author of the book sent to you Could be a textbook could be a book about new technologies. Now you learn you gain a new expertise and It became better. So you perceive the value of the the book that you read To be larger. So the trustee can receive that You see that the equation changes a little bit. I put That letter k and there's a magnification value Next to the remittance that is being sent to the trustee But you can say well, but sometimes That magnification value could be negative or could be zero That's a problem, right? We are hoping that the magnifications be only that it increases the value. We're going to see all those use cases But then You sent the value to the trustee and you see that the size of the pie increased Now you have a dynamics over there where you see the value you can Starting with those points that don't matter, but we're still computing those points, right? You can send in that value. We see the trustee. We see a magnification factor Are you perceiving that the value that you receive is more than what? The trust a trustee sent to you imagine you're paying for a softer subscription Right, you may be paying like ten dollars for it But is it worth to you more than ten dollars? Then You are not positive, right? You're gaining value over there Right now we only see one way. How about if we make this relationship two way? How? Let's get the small percentage of that value that was received by the trustee And send it back to the thruster How can that be? Well, if you Purchase a software or any other product You have to pay for it, right? So that's one way of you sending value back In terms of machine learning you could be interacting with the machine learning model and Sending some information back usability where you you're saying well this result was a good result This one's was not a good result. So that feedback can be used in future versions of of the model to make it better So we're sending we're giving back some of the result. You see that we receive that That g value over there. There is the gain that we received we're multiplying but just like a small factor q That kind of again going to be between zero and one because maybe the trustee doesn't want to send anything back Or maybe it wants and then he sends that value back to the thruster Now the size of the pie of the trustee where it is just a tiny little bit smaller by the value of the pie of the thruster Grew a little bit if the thruster has many trustees Then you have the dynamics where everyone ends up a winner, but it depends on being able to Create value send value that value has to be perceived as A gain of value you're receiving more than you are Giving back and you send some information back in terms of like Either payment for the service or you give feedback on training A machine learning model now you start building a cycle of of trust Let's see some real simulations with that So the first one Imagine that magnification factor that I mentioned before being greater than one And as you remember we started with Magic a million points Right, so we have several iterations on This dynamics in this case of just four But you see over there on the first iteration the trustee is represented by the orange color and the thruster the blue color Right, you see that both of them are gaining from the transaction They're getting more and more points the trustee and And the trust is so it is a good transaction for all of them. You start building that trust And you want to See how this curve behaves over time. That's what I said about the the points not matter But they they do matter Not the absolute value But how they make this curve take shape If over time that curve is becoming steady We're going to take a look more Some more examples later on but if that curve becomes steady You see the trust is being generated And if it starts going down you see some erosion of trust Let's see another use case Where the k the magnification value is equal to one Is the value that you are Generating is being perceived with the same value by the trustee So you see the trustee is still gaining some some value over there Actually, let me go back one one slide Let's take a look at the scale of the the chart I'm changing the scale so we can see the Behavior of the curve changing later. We're going to plot all of them together at the same scale But right now it becomes easy to see the behavior of the curve and how it changes. So this one is going for From 450,000 to 1.8 million The next one you see the scale changed is going from 300,000 to 700,000 So the band is an error the value that is being created is less But in this case over here the trustee is sending something of value The trustee is benefiting from that but the trustee Well, it's not gaining over time Is this all that bad actually no During the phase where you are developing a product for example developing machine learning models And you have early adopters that are working with you. This is a perfectly acceptable scenario Right the trustees that are still working with you. They are gaining value You are gaining value as well because you are debugging the system and You're building in a way where it once it becomes generally available Everyone's going to be gaining from that another case here is Imagine that the Magnification factor is between zero and one so Now you start seeing the case where One the thruster is losing quite a bit And the trustee is gaining almost nothing you see that the value over there Is not really gaining because the scale of the the chart is very small Last but not least is when k is smaller than zero. So there's the magnification factor is actually Reducing value both for the thruster and the trustee In those cases You have a machine learning model that is always giving the wrong results. It's causing The the trustee in that case will be a customer to do real work So you're not really creating value. You are generating more work for For the trustee Not a good case to create To create trust you actually creates a rapid erosion Of trust in that case Let's see everything side by side So you see the scale over there mine again the the scale of each one of the charts They are different, but you see the behavior of the curves Over time they tell you am I generating trust with the trustee or with the customer Or are we losing Losing trust Now let's see the very same picture, but the scale Of the graph is going to be exactly the same So you see how the the effects are only positive when you have the magnification factor bigger than one So the same picture You see the behavior of the curves now the Upper right and the bottom They're almost linear. They're creating pretty much no value. So the the real case where you want to create trust is when your product Creates the creates value that is perceived by the customer that I'm creating I'm receiving or gaining more value than what the The thruster or the vendor is trying to trying to offer. So how do we want to see? That curve going over time of course it would be kind of a Almost impossible for you to have just Achieve a level of trust and stay there forever You're going to have new versions of the product that product can have a little bug here or can have some Issues that you need to address So there's going to be some gentle fluctuation in in trust But you want to see something like this Right, it's a gentle curve. It goes up goes down, but you don't see something that goes Up too much or go down too much the next one This would be a bad A bad example of trying to create trust because you see those high peaks low valleys and the relationship with customer becomes a very Delicate one because at sometimes you're creating trust at sometimes you are eroding trust Inevitably leads at the end. We see over here that customer just Leaves and say hey, we're not going to continue this cycle over here because I'm not getting any trust From the system or from the product that you are doing and even worse you see here that I left the graph below zero When the customer breaks up Because it could be the case that not only they no no longer trust your product but they could go to Other companies or social media and they can talk badly about your product and say hey didn't work for me don't Don't trust this company don't Do this it could be a really negative Value that ends up being with with with customer So this way we can see A framework Where we can measure trust over time Starting from any arbitrary value And by looking at the shape of a curve how it's plotted over time Then we can see are we building trust? Are we trustworthy? Our our customers trusting in us and continue the relationship. So this is one way To do that and with that Let's talk here. Let's talk online It was great to have you all I can also project but I want to understand better I'm not as in the technical weeds and the machine learning side. So I may have some novice questions just to clear So we're looking at here Mathematical propositions and how we can improve the model to create trust on so that when let's say A company is building Their own ai solution They're able to create trust with their using customers to have a predictable flow Rather than that roller coaster graph that you're you propose so Through the different different Like alterations or flavors of their algorithm. These are you're proposing Different solutions Using Using using what I guess I'm confused of If you can explain Explain it like I'm five What's what's going on here? What's the proposed solution and how can companies benefit from something like this? Perfect. Thank you for the question. Let's go back a few slides over here And take a look at this Dynamics again The first one is the trust are creating something of value It's almost like an altruistic act where the trust I can say I'm going to build a product first and Customers will love that And could be a machine learning model could be other products that That there is build but it has to add value. That's the the over there Then I'm going to send that value To our customer. So the value over there the initial value could be the million points that That we discussed the the other slide, right? Then I'm going to be sending it to value to to the customer and We are proposing that the customer has to perceive that value with a magnification factor and that would be that letter K over there, right? So it's going to be the K multiplied by the R. So you have a magnification factor. That's going to be the gain given To the trustee Right now the pie Is grown back and you ask something in return. It's not Necessary but it would be nice to have something in return That something would be a small fraction of the gained value That we are representing with the letter Q So once you receive a little bit of that value back You're going to see that the trustee because it gave something back The size of the pie is going to reduce a little bit And the size of the pie of the thruster is going to grow a little bit, right? So it receives the value and That dynamics is where you can track over time and you want to see that gentle fluctuation on the graphs If the value that is being sent to the trustee Is eroding trust so the trustee's pie is going to start reducing to the point where you just break up And it doesn't matter if the trustee is sending value back or or not Because you're not offering anything of real value to the trustee Does it answer your question? If the argument is that the the K coefficient and the Q coefficient need to be basically, you know Greater than one or whatever it is in whatever context In order to increase trust and to have a more, you know, constant Fluctuation rather than the roller coaster. What are the practical applications in like your actual product design or Machine learning model like training or like in in transparency and explainability like what are the practical things we can do to increase those outputs That's a great question I think you you only going to see that Manifesting over time first you have to pick what are the values that are important to you For example, you can say Is my my server crashing all the time when customers go to the website? Is it crashing all the time? Have I received any reports of unfair bias You can track all those metrics and create a compilation of what do you consider to be value that you are creating Whatever those Metrics that you are choosing you can either assign arbitrary values to them Or some others is are going to be more concrete for example number of crashes that one is easy, right? It only goes up Right, you had one crash two crashes and you can do that. Oh, I'm going to measure that per month So every month you reset the number So you make that collection of metrics And just be consistent On the metrics and how you measure them Then that's going to be your initial million points As I said the number the points really don't matter But you want to pay attention to the shape of the curve Over time Like this one you want to see the shape of the curve over time Irrespective of your metrics Then you see if your curve let's skip a few slides over here So if your curve is Moving like this one where you have gentle fluctuations on trust That's where you see I'm creating value and customers are happy with that I can be affirmative that I'm creating value and being perceived well If it gets like this Good luck Yeah, I can repeat the question over here. I think the question was Okay, please Daniel it's my question. It's been a long day. Oh, yes Because earlier on before the talk even started we were one-on-one talking because By the way doing a talk on the intersection of AI and DEI DEI tomorrow 1155 Sorry, no shame in front of blood But we should work together anyways I'm looking at from the policy perspective. You're looking at it from the mathematical perspective. So definitely Love it, but I wanted to see what is a consistent theme here How do you pull it down to the essence of mitigating bias before it becomes problematic in the AI which then causes problems because that's one of the the root causes of How I can get out of control and start automating automating decisions and making Life difficult for humans when it's excluding parts of the population That's uh, you know these models were traded on so I want to hear more of The implication of that or how when I'm looking at these mathematical models Is the goal to Like yeah, what point the process is bias you mitigate it? It's a great question Yes, bias is always present in uh in machine learning And bias is a qualitative metric. It's not necessarily a quantitative metric However, you can quantify it you can Have processes in place and take notes on Was any case of bias reported? Right, and you can count to the number of cases of bias is reported you can even attach a severity score That could be just a magnifying weighting of the bias. Oh, this was something Trivial this one was something major something critical and each one of them can get that Score and you multiply by one by two by three depending on how you want to amplify Exactly Then you get that What was quality qualitative now you transform is something quantitative Now you can input that as part of the metrics that you are collecting Right, but because it was qualitative you also have to take the The the other side and say Of those issues that were reported in biases How many of them have I issued a fix? For that so if you say there was a bias issue reported And you increase one metric by one And one week later you say hey Uh, the machine learning model was retrained and addressed that issue now you have to Increase the other one You either increase the positive one that the issue has been fixed or you reduce the original one To the original place right you have to Be careful So you don't have this bias metric only growing over time and never reducing because If you issue the fix and the and the reported issue has truly been addressed Now you have to decrease Right because you want to see how your curve goes back It could be A particular case that the reported bias issue causes one of those Valleys that trust is lost But then once you issue the fix trust goes back up Right if you have one valley that shouldn't be a bad thing Right or it could be even a more gentle one if the issue with bias wasn't so bad So there's going to be that fluctuation. So just In your framework Add the number of bias reports that were reported report Bias issues that were reported Attached a magnification factor to it. Is it like a minor major critical? And do the counterbalance once you issue a fix That should report well and fit Well in this framework All right. Thank you everyone. Enjoy the conference