 I would like to know how many of you know the hit HBO show called Game of Thrones? Yes. And how many of you, people who just raised their hand, consider themselves to be a fan of this show? Cool. Seems like the song is very interesting for you. And one more question, is Jon Snow among one of your favorite characters? Yes. And then he's also my favorite character. And if you remember, at the end of season 5, he was betrayed and he was stabbed by his fellow Night Watch members. And then he just, he was just like dying in the snow. And the whole world was wondering if he died or not. And so we're a bit too unique. But instead of waiting for another year until a new season comes out, we decided to predict his death with machine learning. And also along the way the fate of all other over 1400 characters who were still alive at the end of season 5. So our students, they make this method available online and a few days later there was an article coming in the newspaper saying that Jon Snow did not die. And funny enough, this article was found by over 2000 other articles and endless TV and radio interviews. And this was all extremely cool, but the really cool message behind this method is that our students who neither come from Hollywood nor have any experience in storytelling, they excel, they did something really amazing using machine learning. Okay, so what is machine learning? Okay, machine learning, words, they are buzzwords that come from the field of machine learning and we hear them all the time. However, not everyone knows that there is an important difference. So the term artificial intelligence was coined in the 1950s when first machines were developed to perform one particular task as good or maybe a method in which humans can. For example, computers in chess are an example of narrow artificial intelligence. Now, this method, they remained out of the spotlight for quite a while until they became more intelligent during operation of machine learning methods. This method, they were solving the rule-based. While this method, they made observations from data they had available in order to make future predictions about the world. Rule-based means use those rules that we humans impose on the machine. For example, if you see an animal, then check if it's big or small. If it's small, then check if it can climb the tree. If it can climb the tree, then you can see the task. So this is rule-based. This method, they do not in the annual thresholds anymore. For example, spam detection software is a machine learning method. Now, deep learning, this is a subsection of machine learning and it covers the most effective and most complex algorithms we have. They are so effective because now, these algorithms, they can process huge amounts of data. And now, we finally also have those computational resources in order for them to actually be able to process this data. My talk today is about machine learning. So, applications of machine learning can be found anywhere. You can just reach into your pocket where you will find the device that's full of machine learning. If you just want to take a picture, then this device will recognize that there is a face. It will even draw a frame around the face. And when the picture is taken, it will even give a name for the person to this picture. And this is an example of computer vision. Computer vision basically automates everything that we with our eyes can also do. Now, recommendations, you all know that Amazon knows sometimes much better than us if you want to buy it. Facebook also very often knows much better than us when we should be friends with it. So, these are examples of computer vision engines. People quite easily discover fraud by comparing thousands of millions of transactions by buyers and sellers. And if there is something that is not right, it identifies as money laundering. Sales driving cars is a password. We don't really have the meaning yet, but I personally believe that it's quite soon to lose the round. We don't care. This is the area where I, in particular, come from. I'm very excited about this part because it's also full of machine learning. At the university, we were developing methods that were identifying those side effects that are presently one group of people talking together. Because our goal of scientists who work in this field is to develop this kind of medicine that is suited for every single person. And so that there are a few as few as possible side effects that we can live longer and also have here. That is the goal. So, machine learning methods, they are very smart. And now they can do also those things that before humans could do. Classification is one of the concepts by which this algorithm, they learn to map observations, also called pictures, to categories, also called classes. Well, let me give you an example. Imagine you have a data set of cats and dogs. And now we can measure them. For example, for their size, this will be our feature one. We can small or large. We can go ahead and continue measuring. For example, their weight, this will be our feature two, from light to heavy. Now, having these two measures, we can map our cats and dogs into this two-dimensional system. It will look somehow like that. Now, when we train a machine learning method, what we do is that we find that boundary that separates instances of this process from each other. So, that's rather large and heavy. Animals, they found one side of this boundary and rather small and light are found on the other side of the boundary. So, this is training. Now, if we have a new object, we don't know what kind of animal it is. This is called testing. So, what we're doing, we're looking at what side of the boundary we see. And in this example, it's obvious. So, this is on the side of the cat. So, the prediction would be a cat and that would be a correct prediction. So, we see when we train a machine learning method, then we learn from the training data, from the available data we have, in order to make future influences about future objects. And our job of people who supervise and develop this method is that we make as few mistakes as possible. It means that we draw our boundary this way, but not, for example, this way. And this is a very simple example. It happens as good as never that in real world we work with just two dimensional states. Usually it's much, much higher. So, I said this whole story would be about game blocks. So, obviously we do apply machine learning to game of thrones. In order to work with game of thrones, data first we have to get the data. And very good for us. There is a website. There is an Wikipedia and an Encyclopedia of Game of Thrones purely run by fans of the show. That collects all information about the characters. There is a new season that's coming out. There is a new episode, then we send it back to you just at the same time. The end of the new information, but it's full of information. And so we use this webpage in order to collect information from over 2,000 characters of the show. One of the nice things that happened while we were segregating the data, is that all of a sudden we learned new things about the story we thought we would know so well. For example, we found that this show is full of characters. On average, there are around 34 characters in each episode. And in every episode, there are on average about 4 new characters that are always introduced again and again. Finally now, there are even characters for being introduced in another episode, and being cut off in the same episode like this poor guy. And maybe this number of characters also explains why we show you so much expensive. We also saw that there are twice as many men as there are women. However, being a man puts you in a much higher danger, because men tend to die much more often than women do, they're rather safe. Also, it appears that class level does not matter when it comes to the danger, because George Martin, the author of the show, he tends to kill high-born novels just at the same rate as low-born peasants. We also saw that if you reach the age of 16 in the show, that your chances of missing a violent death may also go down immensely. As people after 60, they seem to die of natural death. All of that was extremely interesting, extremely cool, but at the end of the day, we wanted to predict who's going to die. So, for each of the 2,000 characters, all of the 2,000 characters, we collected all information to find him in the city page, such as, for example, what the wolf character appeared in, what house he comes from, culture, what's his gender, age, whether he's married, if his house is alive, and so on. And all together, we found this 24 features could be the most relevant, could be the most important for the production. And now we get to the docs we had, size and weight, here we have this 24 features. For example, you saw that belonging closer to house makes a person die. However, this feature alone is not deterministic. It's always a combination of all these 24 features for a character to die. We used these 24 features in order to make production with a support vector machine, which is an algorithm that performs classification. Here it is shown to be in a two-dimensional space, but it's actually a 24. The dimensional space is just shown in person. So, what happens here is that this algorithm also learns to separate death characters from a life. And if we have a new character who is still alive, then it looks at how similar are his features to those who have already died and to those who are still alive. And if his features are more similar to features of these people, then he's predicted to be dying soon. This is our other work. And all together, we were corrected to 4% of all cases. And this number was complicated by taking all the corrections and dividing them by the total number of characters. And one of the surprising findings we made in the beginning is that we predicted John Snow to die only at 11%, which is like nothing. And we were very happy to see in the beginning of season 6 that he indeed came back from the world of death. Also, our algorithm, and on which in the show her name is Maggie, they both said that this guy he would die. And, yeah, we were very happy to see that he would die. Even though our algorithm had a prediction accuracy of 7% to 5%, it's actually quite high. It was not perfect. And there is this website that our students have created where they put an analysis of which of our predictions came true, which not came true. And, yeah, the last year, the last season 6, it was a lot of fun to observe which of our predictions actually came true, which did not come true, and which for sure did not get control. And to our great surprise and great satisfaction, this show got them my blogging interest by the media. It got covered by over 2,000 articles and with an estimated reach of 1.2 billion people. And I'm saying, oh, this is not only to brag, that as well, but there is also a big lesson to learn from the story, I think. And that is that by... Well, we think that people that connected so much with the story is because it showed that by using any data freely available out there, we can actually do many exciting inferences and predictions. And this idea actually was also confirmed by other articles that said, well, you know what, by using machine learning and this out-of-the-box thinking of our students, actually we can do crazy things. Like, for example, we can even create a new industrial revolution Industry 4.0, where a bunch of students who come from university who don't have any experience with working in industry, just by being that creative, just by doing such cool things with this cool tool can really be great in a field that actually they should not have no experience with. And so, yes, I also would like to encourage every one of you if you have not any experience with machine learning yet, I would like it to start, I would like it to be bold, I would like it to be creative, and also I would like you to change the world using machine learning because it is possible. And I also would like to say a few thank you words to the professor at UNIC who always lived in us, in me, and let us go all kind of crazy things. And to these people that helped with making the story public, also to my co-supervisors who were four people who were running this project, and of course also our students who were actually the true heroes of Game of Thrones at UNIC. And thank you for your attention.