(Applause) Hello. I'm a professor, a researcher,
and a scholar of Computer Science and my name is Hyeyeon Oh. After North Korean leader Kim Jong Un stated that the Nuclear Button is on my desk at all times, President Donald Trump responded "Mine is much bigger, and stronger." by sending tweets on Twitter. Surely the relationship between North Korea
and the America has been improved. But what is important is the button There's the button. For Trump, he has the button of the powerful weapon. However, when we think of the artificial intelligence, have you seen this movie? Matrix. Agent Smith has no button. Even nobody presses the button, it decides and acts by its own will and fights against the people. The important thing is that Artificial Intelligence as well as nuclear weapon could be extremely dangerous. The difference between them is that
the nuclear weapon has a button and it can't do anything if president Trump
or Kim Jong-Eun presses the button. But Artificial Intelligence can be feared since it can do by itself even when we don't control it. So is the subject
that we are going to talking about today. Actually, Artificial Intelligence
doesn't deserve to be feared. Because we can be a master
of Artificial Intelligence. There's no button that belongs to the master. Then what it will be? This is today's key subject. In the house where you live, are you a tenant or a owner? I can see some owners. (Laugh) I'm a tenant. Say I'm a tenant, and something has been broken in the bathroom. But I don't repair it for a long time. Rather, I just leave as it is. Why is that? It's not my property. Regardless of what it is, a house, a car, a computer, whatever else, if it is not my property, and it doesn't belong to me, We become indifferent to it. We don't know exactly how it works. It is because we don't have a sense of ownership. As I told you, let's become a master
of Artificial Intelligence. It means we need to have a profound knowledge
of Artificial Intelligence. That's the key point of today. And If you understand 3 things
that I'm going to tell you in 15 minutes today, you can be the master. The first thing is an objective function, '목적함수' in Korean. And the second is learning. Learning is a prerequisite And the third is generalization and an application
to new fields and problems. But before, we need to know a language of Artificial Intelligence. And actually, we already know the language Artificial Intelligence is a subfield of Computer Science, and an algorithm is the language of Computer Science. The algorithm is Math. You may not like Math. But it is not about a differential, integral calculus,
or something like that It is much easier. Like adding up. Of course it is not the same. But algorithm is also based on the simple mathematics. So, what does it mean that it follows the algorithm? Let's see. It's a bit outdated. You may not know if you are young. It is called Pac-Man. What Pacman does is just roaming around going straight if it has to, if blocked, goes right or rotates 90 degrees counter-clockwise
and goes straight. And taking things in front of it, evading ghosts, and so on. What I just told you is the algorithm. Do some thing under certain condition. Right? Making rules, and abiding it according to circumstances. That's the algorithm. So the algorithm is used for not only acting like this but also judgment. For example, making judgement such as recognizing a car, a human, a face after looking into certain photograph of painting can be done by Algorithm. How can it be done? Is there circle, or oval-like object? And can we find eyes, noses and mouths? After looking for these, 'I found it.' 'It's a human face.' 'There isn't.' 'So it's not a human face.' after looking for circle like object. But in most of cases, there are so many variables. Every person has unique appearance. Lights can be either bright or dim. A photograph can be taken in the distance or closely. Many things need to be considered. Because of these, With a simple algorithm as I mentioned just before is almost impossible to tell faces
in the photograph. So, Now we have finished about the algorithm. Artificial Intelligence is needed for the better judgement that the simple algorithm. That's the conclusion. So the first is objective function when I said that we need to be the master
of Artificial Intelligence. If you want Artificial Intelligence to play Pacman on behalf of a human with a keyboard or a mouse, you can just done it by defining objective function to Artificial Intelligence. Your mission is to get a perfect score in this game. Then what will Artificial Intelligence do? It playes by itself, 'I tried but I didn't get perfect score, I was killed.' 'I was caught by a ghost.' 'Well, then give it another shot.' Then it gets higher scores, improving skills until it gets the perfect score, and finally, "I got it!" so it learns algorithm by itself. The objective function is the most important thing. Because I need to get perfect score
in this game. If the objective function is to get 100 points in the game, it will quit after get 100 points. It will do nothing after the achievement. So the objective function is the important element that can change a result of Artificial Intelligence. So, because we need to be the master, I'll make it comply to my order. If that's what you want, one of solutions is
to change objective function. To take an example of an autonomous car, an autonomous car has an Artificial Intelligence algorithm that recognizes pedestrians. It should have,
otherwise it will hit pedestrians. But say, that the data that is used for machine learning to recognize pedestrians consist of 999 adults in 1000 data and only one child. And if we define the objective function to recognize only the pedestrians that occupies the majority, what will happen? It may not recognize children. It shouldn't be like that. We need to protect the vulnerable like children, shouldn't we? Since it is an important matter what do we have to do?
We need to modify the objective function. For example, If the machine doesn't recognize children, or recognizes children, plus 900 points. One or two points for adults. Then it will make an effort to recognize a single child. In that way I mentioned , we need think what kind of the objective function should Artificial Intelligence have. Thereafter we become
the master of Artificial Intelligence, having it under control. This is our recent research
of the objective function. I brought it here to show off the results. Artificial Intelligence does
lots of interesting things, actually. What we studied is about the relationship
between a language and a color. When we think of the word 'dream', we can associate certain colors, right? As we can see, there are 'Children' and 'Teen', each has different colors. Children is a pastel tone,
such as sky blue or yellow-green Teen is a kind of blue color,
but it is a bit darker. This is found out by modelling a correlation between contents
and paintings of the cover of the book sold in the online bookstore in the America with Neural Network
that is used in Deep Learning. So with that out of the way, what is the objective function here? It is to predict the distinct vision of the color well about the explanation written in the language. And the second, we need to learn something in order to become
the master of Artificial Intelligence. We need to learn. We have the objective function. However we need to learn from a set of data
from that function. For instance, this is a kid,
around 3 years old. He or she is getting to know about the world, asking about animals. So a cat, a dog, mom says "that's a cat." "that's a dog.", and the child learns. But if mom is fluent in English, or the child lives
in the English speaking country. Then instead of saying in Korean, that's a cat, That's a puppy, mom will tell the child in English. And the child learns English. Artificial Intelligence is same with the child. If you give it a data in English,
it learns English. If in Korean, it learns Korean instead. We study Natural Language Processing
like things that I just mentioned. So take a look, words that are in similar category are grouped.
[Fruit] [Animal] [Fruit] Cat, dog and names of flowers. If you give it numerous sentences
including words like these as a learning material, it learns very quickly and easily. Artificial Intelligence in this level is common these days. It can learn any languages that exists in the world
if only learning material is given. After learning with model
that has indigenous characteristics of Korean, you can see conjugations of verb - let's go, to go, went,
[Future] [Present] [Past] let's play, to play, played
[Future] [Present] [Past] let's meet, to meet, met
[Future] [Present] [Past] - it's quite different from English. English does have verb conjugations. But own characteristics of Hangul,
its character consists of initial, medial, and final consonant. The final consonants changes continuously, as you can see. So to make Artificial Intelligence like this
that can learn the rule, it needs not only the data, but also profound knowledge of Korean. This is because
when we teach Artificial Intelligence, a consequence of learning depends on which data is given,
and how do we put the characteristics of the data. Third is the last what we need to know
to become the master of the Artificial Intelligence is a generalization. What is the generalization? We generalize too often. So what does it mean?
Say, think of the child, that have learned about a dog just before. The child realizes, "Each dog has
different sizes, colors, and ear shapes!" Now he or she learned how dogs look like. And after the child went to a zoo and saw a cheetah.
Then 'there's a cheetah!' 'Bear!' The child saw a bear. The child already learned that each bear have different sizes and colors
even he or she saw it only once, from the empirical knowledge of the dog. In such wise, applying the knowledge and empirical evidence into a new problem or a new field is called the generalization. This is the third important key concept of Artificial Intelligence. Here are some interesting results about the generalization. This is a landscape photography. And if we apply this photograph to a new model, What will result from that? For example, if we make models by teaching a style of Van Gogh, and then put the photograph into that model, then the result comes out like this This is the one with a style of Edward Munch. This is of Vasili Kandinsky. So after building a model, and applying knowledge and something else
that I already have such as abilities, skill to new things is the third core element of Artificial Intelligence. and a method to control
Artificial Intelligence as the master. So we learned 3 things today. To be the master of Artificial Intelligence,
you need to know these three things. And be cognizant that you can change these three elements. If you change these, you can control Artificial Intelligence
as you wish. This is everything I have talked so far. If you stand back and think,
you'll realize it quite interesting. Because those three things are methods
not only to become the master of Artificial Intelligence, but also compatible with how to be a master of your own life. In our lives, what is the purpose of my life? It's like thinking of the objective function. And to achieve the objective, what should I study experience and think to learn? And lastly, knowledge I have learned, experiences that I have, abilities that I have, with that things, where or which new problem or field is applicable? So it's time to think of these three things
that you've heard and the future that becomes the master of both your life
and Artificial Intelligence. Thank you. *Applause*
