 Let's take a look at some of the things that I saw in the video about handling things. First, we made a new step, so because of that, so like I was mentioning that our break-up was not going to be two parts. The first is, this is a Tokyo planning machine learning event and what we are going to do is do a Tokyo planning machine learning or Tokyo planning machine learning as a case and the second part is where the planning is going to speak about music and cases on patients who are going to be able to do it Just to give you some background for back to the release of South Side it's going to come that health officers they are going to be able to participate in classes So we power up that automation We run the work of the office of Installer for 60 billion hours a day so 60 billion and 20% of us do it a day We would then send these based through protection fires, organization fire, etc and it would be 10,000 to 10,000 fires in the fire room any if I could come to the institution fire for example, I have a question and I have that from these films of 75 and just because 10,000 we have around 410,000 we have to pay that 10,000 a day This is a scale that we have to do here by machine learning So So when we are busy, it's like what? Conflicts, data? So this is our answer question Data, yes It's actually from the TV that it's slightly older than what we wanted to do here on the first of the day actually, right? Our play is at least 200 Let's say our 100 is the number of questions The number of questions Yes, it is the standardization so 60 billion hours a day and then each grid so each idea person corresponds to grid, so people are going to think that the stock exchange is very you know, I mean, somebody comes in and says this is the actor, I must go and then we can talk about the fires that are going to take us to the institution So each of that is not a transaction So you can come to the institution fire So some of that people if it's going to be there then it falls to all of the fires then how much are you going to spend a green person based on your position and because that's the type of data so that's one, two Let me tell you what I'm going to say and I'm going to say 50 what do I say 312 the question you were in the 300 years wow impression sorry, what do you mean? grid grid and each of these grids has a lot of information like what do we buy and what grid are we going to see this is counter-couple who is going to be counter-coupled and all that stuff so all of that is a transaction for us as you can imagine all of this is 118 billion that's great and as you can see the impressions this is the number of grids per second there can only be one winner for every single person so you can sell and add one of the grids on that particular which is the number of grids per second and this means that you are running master models of infrastructure think about the connection that we established each of these need to be processed in 20 milliseconds that's what that's the time that we have for the first stage of our transaction and there are other parameters to evaluate in front of it and all of those that happen so one of the problems that we have to case is actually infrastructure costs one has to ask but also the way that we're sitting out for the partners that is you don't want to process and they need to originate so you have to sit all the way to the place where you're sitting so one of the things that is very alive and healthy is to learn and predict grid fires are likely to be done in the future and who are most likely to win and make their decisions so what that has to do is that the fires are a very particular thing and in fire we do some fires that we use in school specific audiences others we do use in public and in somebody we do use for specific the question is kind of specific how can you do that that's where they are the business it's up to you to measure and adjust your location so every buyer is different in what they are looking for and you're just going to look at what's the setting of the grid what's your interest for particular fire or what's the unit of access there is a question and it's a very specific And one of the matters by including the initials of this is the efficiency for all people who are listening to those requests from us, so by us, as well as the towards the cost of infrastructure that is needed to support that. And in our global example, we can see that it comes very soon to have a request for a second that is sent out to the partner, which will improve quickly, because now we can improve and produce more projects every month. And if he goes spend money for PTSD, he gives space for second hour. So, you see that the supply of energy on the air is very common and it's not really like that. So, how much of that is in general? One of the issues by state is the very organization. So, you are talking about this, it is very granular to the inflation rate. What is the value of inflation coming from the value of the government, how much So, some of the variables could be like time idealization. If you think you see traits that are depending below versus traits that are depending on the end, that are going to be much higher than the end. It's a perfect time. These are specifically higher. In the same case, they are basically higher than the end versus the end. There could be different demand aspects. There could be higher than the end. Some of those are going to be more common than others. But in that case, the size of the properties are going to be more valid. For example, if it's a really new step up, the value of the ingredients there are going to be much higher than the end. For example, if somebody is going to get a system block, that is not going to be possible. So, there are a lot of factors that are coming to figure out and that can affect the price of an impression. And we are also learning specifically to figure out how the right price for an impression should be. And that should increase the publisher's rating. And it has improved here. And we have also observed that the market is not necessarily more replying to various industries. And the impression is of course to carry out the report card. We are also going to just go and buy a product that's there. But because of the price of an impression, carry it in and keep it as it is. I have a question. Sir, I would say that the strange survey of just one model is a product that is very useful. So, it's not stupid today. I guess that's the subject that you are asking. They are looking to kind of reduce the price of an impression. As in, the models are very useful, right? And then, depending on your time, I would suggest this. The 15 years that you are going to be, that actually look at the model, the quality of the product, the uses of the model, the ingredients. All the 20 years that you are going to be, you are also thinking about the system that you are going to have. Actually, I raise every single impression, every single, we observe 60 million impressions that you see. The model that you created, every single model that you created, every single model that you created. What is the second thing? Second thing? I'm not going to mention the third one. All right. I think it's a case. So, we covered all these cases. One of them. Like, one of them, what kind of connection problem one was. In fact, I talked to my client that I know a lot about predicting what kind of impression it would be. How do they come into these cases? And how? Secondly, because of our interest. And do you have a like, what? So, I think secondly, since this is the EP file, there's interest in putting more specific time points in there, depending on what it is. I'm not going to show up at the company, I don't have any interest. Each of the users are interested in both. If I am, let's say I'm a reader, every user is looking for a house. I have a house, I have a house, every user is going to buy a car. And each every kind is interested. We can have a lot of signals, not that interesting, not that user. So, what the user is inside of the city, what time is it taking, what's the point of time on the page, all of those kind of going through. So, what is the user? So, what is the user? I think interest in it. And what the user is looking for, they can purchase it, they can take it in the case of that. And that, on the file that comes out, the user can come to the user's and to know how long it's going to take. You want to see more of that, you want to see and that, but I don't want to review that, I don't care about that. So, this kind of personalizing I think we add, I think it's a good thing. And those kind of buttons are very similar to the sense that buyers who want to see the customer that is the kind of user that I'm interested in going on that. But you may buy some, but they don't have these audiences that are coming very, very far from that. This will be a very small number. And these models can be used for other things, user, type of users or audiences. So, I use this as a scale in a model to come up with a look-alike product and increase in the scale in which these impressions are made out of the user's mind. So, I take today down to be look-alike models that are able to be used in order to see the market as users. And using look-alike models that you should buy, and if I'm more user, I'm more faster than this and I'm not more interested, but here it's a very similar kind of a stitch to the electronic user or the customer that I have. And then I choose to increase the scale of the market. That's similar type of users to the life that I've been doing. And there are some people which have taken a different approach because I actually have applied and in fact, who are just making full users. Like, how to use the similar type of users that I have seen today. And sometimes, it's not very intuitive. So, for example, let's say I'm a opposing company and I'm looking to start looking to reach out to users who may be a complex customer. Or I don't think it's easy to look at how much money I actually possess on us because they are the ones who typically come and pay to the side of my business name and increase the prices. But what you can see is that it's not a purchase or an opportunity to use it. It's not actually business orders that I've been doing. It's just the data that I have and the other types of those. So that's a point that I've been trying to support everything. So, in fact, it's just making actually the same business to us or the business to us. But I can actually do the same way that I do now. They're the ones who are really a customer. So, my name is and this is Peter and this is my business name. You may not really know what it is. It's really, really easy. In addition, in these cases, we use machine learning and in the team not for that hard time. Also, kind of, as you mentioned, it's a very easy to use and not to do any sort of simple work. And the same comes from a campaign company in the past seven days they don't come in and they don't know what to do and so on. They just come in and they don't know what to do. So, in my talk, I would like to talk about the campaign in the last seven days from the United States of America. And this will get processed and used in the exact work that you are going to do. But if we can have an opportunity to use the fine information that you need to access, now, no more information like women and non-women traffic. For both of you are sensitive traffic because there are a lot of thoughts about traffic and related by not as non-women traffic but she is making traffic just to monetize the traffic. So this traffic is coming to women for non-women traffic. Because I also want to say about non-women when I said how many impressions is, how many impressions is the question that we discussed in this case is to figure out how many impressions are there for the campaign or how many expressions can be said So that we can keep on doing that in the buying and selling products. Similarly, we are making a plan to figure out what is the quality of the users which we are going to be doing in the future. And also, we are actually going to do the buying and selling products. We are already in the selling and selling products. The next step is, we are going to be using machine learning for inclusive processing at Skate. That's what the infrastructure is talking about. This is a problem including an existing skill and learning system. The next thing is that you are not going to be using for a publisher. So you have to look at the main problem. What is the efficient way to buy a specialty? You can use it for databases. The next thing is that you can use it for many different traffic. For passing, you can figure out how is the candidate going to deliver So today, this part I am going to cover about the musings on cases of genetic, so I assume I assume that most of us here are trying to show you step 1 through the cases of genetic tests. So here I am going to share about what type of musings or whatever it is of course we have on the cases of genetic tests which we have in the afters of 999, so the key parts here are we should, when we take the requirements we should always design the final score in order to make it interesting. So what I think is one of my favorite is we should come up with a final expectation saying like I should be able to recognize the percent of users who are going to give my ad in next 2 months. So and with the accuracy of something I can have good thing as a bit of this I am going to take. So the second thing is this is happening maximum by 15, if the classifier is at all this, you will have that much of minimum accuracy so that is the lower volume. We can also get an upper volume so the details which we have, even if we include all the variables which are available, still if you find the output rate by which that is the highest rate the details which we have, so that is the upper volume, so what we use or what level we can release it to, you will have to reach only that. So knowing that is quite an important step and if the machine learning project already defined some other things. It is always itself the domain. So the other data scientist sometimes you can get the data unless you understand like for example and domain you won't understand any of the charges or the things which we do. For example when he talked about bit problem, so if you feel or if you want to know what are these things you need to know it, you know what is increased my revenue, increased my revenue or reserve. So that is something we only get to know unless we understand the domain in depth. For example remarking is a very general term but it has a very specific meaning in that it means like it's only the user who has already been to our website and hasn't been installed on that browser. So that is the only who we can remark it or we can check. So the definition, general definition we want to change is the context that we have in the domain. We must always take care of searching the domain. And the key fact is this correlation is a very important metric and we should always go into it. Like he said this something is affecting my, I can insert or something but we should always be aware of a system scenario. What this paradox is is that if I am comparing two things the correlation will only show me the current results considering all other things are seen. For example I am comparing two users in an app I should compare like all other things we can see. So there is a proper, there is a very famous example which happened in UC Berkeley and now it say they were buying sugar spilling water which in quotes meant it has a bad component. But this was done with the basic stats then how many students were getting used to it. But I understand that they actually did a comparison and department wise they actually found that which actually got affected department wise is somewhere to this. So that being the paradox here was that two departments which were very competitive and in which very few people did it. So that was the basic paradox. So if we always correlate what we should understand like I will be correlating it. The next again not there is a question always should be some descriptive line. So as an engineer sort of data we always want to get something good out there and then we show to the customer. But this is the wrong thing and we should always show for our initial descriptive analysis which we do on our data to the kind so that we understand what we have learned is many sense in the domain or we can ask domain expert like what we are correlating or what we are understanding from the script there is actually many sense in the context. So this is very key thing unless we go on this if we go wrong on this we might be building a critical model but it should be discussed. So again the examples of how we should always we also have to aggregate our preferred data. Since we need for example an act as we said we need to see 60 billion impressions but each impression is independent. So you can't do that machine learning already because it has very little sense in it also. So you have to aggregate your preferred or something like for example if you are interested in pivoting a new prediction, you can pivot a user and then get all those characteristics or you may want to pivot once on that site for the publisher or advertiser and then what is it possible and then you will be able to see learning from that. So as you can see there are so many like there will be dramatically more than 500 different peaks. So each of the different peaks and different meanings and some of them proxy for each other, some of them are not appropriate also. So what is important once you are able to do data analysis and after collecting activities and you do understand it if you might want to even faster than they have to understand which all the peaks are related and which can proxy for other and they include fewer features so that they have lesser features to possess in the final model as compared to if you are just taking the number of all the peaks so the important thing when you attack is most of the peaks are not brought up for all the cases for example the mobile app ID is the location and there are various reasons for that but when we get data it's like what is how much data we are getting for example if you are getting mobile app details or location it's only, I am also talking about environment model using location because you can't do anything about it so the other option is to there are cases where we can include data of IP and other things so we would have to obey to that to do that this is for when we get to advertisers modeling so I would strongly recommend that we should always build the business oriented advertisers before we can move this is important because building advertisers would help us to get the understanding of the business and we should also review these things at any time so that we understand that we are not doing anything easy here we are not doing wrong which is not the perfect sort of business we are so also when choosing ordinary company we always go ahead and say which is the best model like for example if I have a pacification app and I have a pacification so be there I will go for it but this doesn't always work for example in a company when we had this we should be trying to understand but the one we found was this company was giving us a similar result as a brand of pacification with the difference of existing of the company that was the big problem so always considering the requirement of how we are to implement this model in production so to build the models we have to then obviously compare and you should always as I said to know always when to stop and when to go ahead if it is after the we always have the problem of doing the category of how we are going to produce this wheel the future can have so that is again a big problem considering millenium pressure we always have so many things to come up so instead of going to have a lot of approximation which gives us better idea of what the time is but later in the middle period it is short time so rather than using complex we always have approximated we know when to stop considering the rate of is doing one thing is worth the effort I am okay with adding question of but still it is what is this we should know there and stop there so we have this I think the most important part of machine learning is presenting the model even like Ajay Mahal said Mr.Hauske company is very important there is a problem as you know you should not create a business of model models are being communicated in a simplistic way and which is our vision being a strategy that happens and we will try to use some of the domain language so that the business or the end users we are presenting to better connect with the presentation and they can better understand for example like chat is always the best thing and we obviously use different games ads so what is different games ads is like how much game you are getting by exposing it to how much it says you can add a different ad aid with 8% of customers we are likely to convert it by 2-20% of the population so this is what the model is the lacklines was the base name it can handle how it will do and the game shows you like how much benefit you have the management always ask for that how much revenue you have what kind of model we show the first question versus how much revenue you will be getting from it but how much revenue you are saving from so this is where you just connect in your line and say we are going to reach not only 8% of people but similarly just 20% of people so this is the way you can connect with the business so please let the people know about deployment and operation so sometimes you know flash models don't get a good product so as we know we have like 200 seconds to complete a real-time reading process so there is a percentage where you send it to to the predictor and finally it shows their results so by the time you log in to your Google or type of media you see a lot of changes so this is what we are talking about in under 200 milliseconds so if we are considering model printing here or applying machine learning even the predict method cannot take less than 10 milliseconds for us so that is being understood so we have to consider the environment statistics like what is the output format of the data structure so these are the critical requirements and even the example in Amazon even the second page is dealing because in Amazon I don't like dealing covers so this is an example of how models are critical if your model isn't printing to the environment better then it makes very critical sense so printing activity is obvious even if we print a model which is giving all the accuracy now it may not because the options are not the same and at the options are not the same the people buying habits change the season change happens so we have to always keep on the model up to date and up to date all the machine learning is again and again for example the best example is an etiquette example they ordered a booklet but they could not buy the production the reason was when the competition started it was a video editing site which was printing beauty they are going to go online they have to sell out the money but all this could not be different because they just came to teach you these are the key points which are the important if you go in the middle of the model the important is how could we define how the option is so yeah I think let's start we just have to chat I am going to start I am going to start please we are going to do very very first question I would like to know do you have any experience with this style of writing of working so what you are telling me what my character has performed do you feel like is this the story you are taking So, scoring is, if you say my model is scoring this much percent and you all get actually what should be the ideal percentage, what has changed? As I said before, if we have base rate, what should be the ideal should be? For example, I got a 2, if I get this amount, it should be by now 80-90 percent, but my scoring says my current model is about 60 percent. So, scoring is just what a model has, it is what the valuation would say, actually is a model called now, you know, scoring is getting what the current thing is, scoring is just getting the valuation, for example, now model is more than 80 percent now. Valuation test, this is 80 percent. What is the, like, I have been in model side, I have been in model side. So, if this is the case, I am, you know, and you don't usually get the rate of the score in the model, so you can score some of it on a traditional data basis, but the levels will be there in a few seconds or two in a few seconds and we will have. So, all of the techniques, you can score different, and we will, we have got a lot more questions than we can, you know, a lot more questions than we can, we have got a lot more answers than we can. So, when you get a task, you get a few seconds, then you can have a few, a lot of things like lowest level, lowest level, lowest level, lowest level, that can be, you know, all that stuff, it can be, you know. Any understanding of great chocolate in, you know, so, what, like, for every model to be built in one, how do you go ahead and adapt it? So, when you adapt it, yes. It's a pretty much the same model, when you adapt it to one of the teachers, who can send it on one. So, you can get the single, you can get the model, then you can do it by the, you know, figure it out, and then the single, you can send it to my teacher like this. Okay. Yes. Can we go like this? Okay. These are extremely important people in the, you know, to build a particular pilot, and then any pilot can find a partner. So, then you know, that it's going to be a very good pilot, it has a lot of value. So, working as a pilot, you can really do a lot of work, and raise a lot of money as a total in this case. So, then again, you know, I think it's going to be a good one for an app. Probably a good one for an app. It's just really good. So, that's it. Many people, these are extremely important people, so, that is important. Device life is going to matter a lot. So, devices in the long, in the long run, are coming from, you know, higher standards, you know, so some signals that you don't want to to buy a new service. So, any kind of money that you don't want to buy has to be some serious, very estimated, but I understand what you're saying. So, if you buy a pilot, and you're a very good person, why don't you buy it? So, that's going to be good. And this pilot, it's going to be very good. Yes, exactly. This one is going to be good, but it's going to be very good. So, if you have any questions, please let me know. If you have any questions, please leave them to me. We'll be very kinder. We'll be very kinder. But, it's not going to be good. It's going to be very complex. It's going to be very complex. It's going to be very complex. Always, you have to use BNM on the device that you're going to be in. And, everybody, just a second. If you have any questions, please leave them in the comments. Thank you. Thank you very much. Thank you. Thank you. Thank you very much. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.