Hello, I am Seok Jun from Korea University, responsible for the lecture on data learning and intelligence. I am pleased to meet you through this video. In this lecture, I will primarily discuss an overview of artificial intelligence. The lecture will be divided into three main topics. First, I will explain what artificial intelligence is and the process through which the AI we see today has developed. I will discuss the evolution of artificial intelligence. Next, I will cover the most critical technology among the artificial intelligence we currently use, which is machine learning. Therefore, the title of this lecture is data learning; the learning here refers to machine learning. Data learning and intelligence—intelligence refers to artificial intelligence. So, I will provide an overview of machine learning. Additionally, we will look at how these concepts are applied in practice.

In data learning or machine learning, programming is indeed a very important element. So, lastly, I would like to discuss how we can approach programming. Of course, programming varies widely from basic to advanced levels, but here I will mainly focus on the most fundamental level, helping you understand programming and providing you with the simplest tasks you can actually try.

First, let me talk about the development of artificial intelligence. We often refer to it as AI, and in English, it is called artificial intelligence. As you know well, artificial intelligence is a topic that is being discussed everywhere in our current society. For instance, we use AI to predict various situations or to identify when a person is in a dangerous situation through CCTV.

Artificial intelligence is being applied in various fields, and if you recall, it has only recently become an important issue in our society. The concept of artificial intelligence has existed for a long time, but its direct impact on our lives has not been as long. In fact, the most widespread use of artificial intelligence in our society began in 2011, when IBM developed an AI called Watson. This AI was designed to answer questions posed in human language by searching vast dictionaries or various documents to provide the correct answers.

So, the company IBM initially developed Watson to apply it in various fields. To showcase Watson's capabilities to the general public, we entered it into a quiz show called Jeopardy. A quiz show is about asking questions in human language, like, Who invented Hangul in the 1400s? The answer would be King Sejong, which is information that can be found on the internet. Since it's a quiz show, you can ask questions and provide answers based on information from encyclopedias, making it an excellent way to validate Watson's abilities. In 2011, Watson participated in the American quiz show known as Jeopardy, as shown on the screen here.

I appeared on a quiz show, and here are two human champions. This Jeopardy quiz show features people continuously solving quizzes, and these are the human champions who participated. Together with these two champions, we are solving quizzes, and as you can see here in the store, the prize money from IBM in the center is the highest. What that means is that we have defeated the human champions, thereby significantly demonstrating the potential of artificial intelligence. Since 2011, after Watson, the spread of artificial intelligence in South Korea was greatly influenced by an event involving AlphaGo, which had the most significant impact. This happened in 2016. Until then, the game of Go, for example, had existed alongside chess, but chess had...

There was a time when computers could defeat human champions. So, I thought that computers, particularly in chess, would be better than humans. However, until AlphaGo, many people believed that artificial intelligence could not defeat humans in Go, as it is a very complex game with complicated strategies. But in 2016, during the match against AlphaGo, as you all know, AlphaGo defeated Lee Sedol, who was considered the best Go player at that time. This victory made people realize that artificial intelligence can accomplish many tasks and can perform certain activities far better than humans. As a result, interest in artificial intelligence surged, and it became a pivotal moment for the widespread adoption of AI, especially in our country.

Then, for a while, artificial intelligence became very prevalent in society and was being applied in various fields. Recently, with the emergence of ChatGPT, artificial intelligence is now capable of conversing in a manner similar to humans. This has led me to think that artificial intelligence has truly come very close to resembling human interaction, and we may have reached a point where we can no longer distinguish between artificial intelligence and humans. As a result, artificial intelligence has become widely integrated into our society. 

To understand how this artificial intelligence began, we should first look at the concept itself. In fact, the idea of artificial intelligence has been around for quite a long time. It is not an exaggeration to say that the concept of artificial intelligence emerged simultaneously with the advent of computers. Historically, humans have always compared themselves to animals or other entities.

The differentiating factor is that humans can think; humans are beings capable of thought, which has been quite different. However, upon reflecting on this, I began to wonder if it would be possible to create a machine that could think like a human. This concept of how to create a thinking machine was first introduced by the British mathematician Alan Turing. When we look at who Alan Turing is, he is often referred to as the father of computer science and is the inventor of the modern computing machine known as the Turing machine. Of course, before the invention of computers, there were calculators. When we consider what a calculator is, we see that it is used to perform basic arithmetic operations.

Such arithmetic operations can be performed manually by a person, and they can calculate directly. However, even though this is a very simple task, calculating with large numbers can actually be quite difficult. Therefore, the ability to perform these calculations quickly with a mechanical device led to the invention of the calculator, which was already developed in the 1640s by various individuals, including mathematicians like Pascal and many others. However, beyond the arithmetic operations we think of, the question of how to create a machine that can perform logical actions, which we refer to as algorithms, was a very challenging problem. This was the initial idea about computers that Turing first conceived. The person in this photo is Turing himself, and you all...

You might like this, but there is a movie that depicts the life of Turing, and that movie is called The Imitation Game. Here is the poster for it. The actor you see is well-known; he plays Doctor Strange in the Avengers. If you are interested, watching The Imitation Game will give you insight into Alan Turing's life. 

Now, the Turing machine that Turing developed is a mathematical concept, but it can actually be constructed. When you build it, it looks like this. I won't go into detail about the Turing machine in this video. 

Ultimately, what artificial intelligence means is that humans engage in various logical actions. For example, in different situations...

In terms of recognition or cognition, we make appropriate judgments based on that cognition. All of these actions fall under logical behavior, and the machines that mimic such logical behavior can be considered what we think of as artificial intelligence. So, we want to implement artificial intelligence; we want to create it. However, while we understand what artificial intelligence does, which is to make judgments like humans do, it is not clear how we can actually create it. In fact, we do not know the method. For example, there are several intellectual and logical problems we need to solve, such as calculating the trajectory of a rocket when it is launched. Calculating the trajectory of a rocket involves using various physical laws expressed in differential equations, and then we...

계속 풀어나가면 결국에는 궤적을 계산할 수 있습니다. 이건 우리가 어떻게 푸는지는 잘 알아요. 하지만 이걸 일일이 인간이 개선하려면 굉장히 어렵죠. 그럼 이건 어떻게 우리가 풀 수 있는가? 이러한 것들은 어떤 프로그램을 통해서, 컴퓨터를 통해서 이제 우리가 어렵지 않게 구현할 수 있습니다. 하지만 얼굴 인식 같은 경우도 우리가 이제 논리적으로 어떤 판단을 내리는 건데, 지능적으로 판단을 내리는 것입니다. 이 사람은 내가 아는 사람이냐, 이 사람은 내가 모르는 사람이냐, 이 사람은 나의 부모님이냐, 이 사람은 나의 친구냐, 이 사람은 내가 모르는 사람이냐, 이런 식으로 우리가 어떤 판단을 내리고 싶은데 이러한 판단은 사실 5살짜리 어린아이도 할 수 있지만, 컴퓨터가 이걸 판단하게 하는 것은 굉장히 어려운 문제였죠. 그래서 사실 우리...

The idea that a computer could perform all logical actions that humans do has been a challenging problem for a very long time. People have tried various approaches, but they often found themselves at a loss. Eventually, the idea emerged to create something called artificial neural networks. So, what are artificial neural networks? Essentially, all logical actions we perform are carried out by our brains. Although we may not fully understand how we do it, we can study the way our brains function and mathematically model that process. By having a computer execute this mathematical model, we thought it might be possible for the computer to effectively perform a wide range of tasks that humans do.

The human brain exists here, so when we ask neuroscientists or biologists how the human brain is structured, we learn that the human brain is made up of many interconnected nerve cells, called neurons. Then, we ask further, How do these nerve cells function? Biologists explain that nerve cells receive some kind of stimulus from one side, which leads to internal chemical or biological processes that send signals out from the other side. This process is essentially the input that activates the nerve cell, allowing it to send out signals that help us transition to the next stage of communication.

이제 인공적으로 만든 신경망이 바로 인공 신경망이라고 부릅니다. 이러한 인공 신경망은 실제로 우리의 뇌의 구조와 기능을 모방하려고 시도한 것입니다. 각 퍼셉트론 노드는 입력을 받아들이고, 이를 처리한 후에 출력으로 변환합니다. 이 과정에서 여러 층의 퍼셉트론이 연결되어 더욱 복잡한 패턴을 인식할 수 있게 됩니다. 따라서 인공 신경망은 데이터에서 학습하고, 예측하며, 다양한 문제를 해결하는 데 매우 유용한 도구가 됩니다.

Because it is artificial, we attached the term artificial. Thus, we developed a mathematical model known as an artificial neural network. We thought about what we could actually create using these artificial neural networks. However, when we tried to implement this, we realized that while there are models available, it is still unclear how we can continuously train these models and how we can further develop them. People have been pondering over how we can create effective solutions and have attempted various approaches. Consequently, they discovered two major, distinct aspects.

I started to think that we might now be able to create artificial intelligence using an approach. So, these two types are rule-based artificial intelligence and data-driven artificial intelligence. I wondered if we could create artificial intelligence that is not possible with these methods. Now, what is rule-based artificial intelligence? We often refer to rule-based artificial intelligence in English as rule-based or sometimes as knowledge-based. Essentially, it means that there are clear rules in place. Based on those rules, the artificial intelligence makes certain judgments. In fact, many of the judgments humans make are also quite rule-based. For example, when you go to a bank to apply for a loan...

Or, if someone is trying to apply for a scholarship, the person in charge will determine whether the applicant is eligible for a loan or a scholarship. They will have to decide whether to actually grant the scholarship. So, how do they make this judgment? Typically, there are guidelines or rules that are referred to as criteria. For example, they consider the person's income, and if it’s a scholarship, they look at the student's previous year's grades. They assess these factors, and if the grades are too low, then the scholarship will not be granted. Alternatively, if the student's family situation is relatively affluent, then they may not qualify for assistance. In this way, they evaluate and make a decision on whether to award the scholarship to the student or not.

there will be rules for such judgments. Then, according to these rules, we can simply make decisions. If we give a computer a set of rules and instruct it to make judgments based on those rules, this becomes rule-based artificial intelligence. Such rule-based AI can actually be used in many areas. For example, in the case of autonomous vehicles, there are rules such as stopping when there is a person in front of the car. Right? Such rules start from the most basic traffic regulations, like stopping at a red light and going at a green light. For instance, if there is a person, the car will prioritize stopping. We can implement a variety of scenarios based on rules, including those that involve value judgments. So,

In rule-based artificial intelligence, the most important aspect is to collect and organize rules from various sources, which is actually quite challenging. For example, in cases like loans or scholarships, there are clear guidelines or rules to follow, but in areas like law, there are numerous rules that are intertwined and often contradictory. In such cases, how to resolve these issues is indeed very difficult. Therefore, rule-based AI fundamentally exists, and one example of this is the Clinical Decision Support System, which helps us make medical judgments. For instance, it assists in determining whether a person has a cold or whether a specific medication should be prescribed.

There are rules for judgment that exist, and we can use these rules to employ rule-based artificial intelligence. However, upon closer examination, we find that there are many shortcomings with rule-based AI. There are certain elements of judgment that cannot be addressed through rules alone. In such cases, we need to utilize data-driven artificial intelligence. For example, a case of implementing data-driven AI is facial recognition. When determining whether a person is male or female based on their face, it is quite difficult to think that there are clear rules distinguishing men from women. However, while there may not be explicit rules, people have their own approximate rules, and based on these rules, we can make judgments. So, how do we establish our own rules?

When we think about whether we created it or how we should think about it, we realize that from a young age, we naturally learn the concept of gender through interactions with our parents or friends. We come to understand that certain types of people are called boys and others are called girls. These concepts are something we acquire naturally from a young age. So, we see that some people are boys and others are girls, and this is something that someone tells us. They say, These people are boys, and These people look like girls. These observations become data. In a similar way, we give a computer a million pictures of boys and a million pictures of girls, and then we ask it to find the rules to distinguish between boys and girls based on that data, and to make judgments accordingly.

What we are talking about is data-driven artificial intelligence. Data-driven AI means that computers must naturally find and learn the hidden rules of judgment from data, which is why we call it learning. This leads us to the concept of machine learning. The most important aspect of data-driven AI is learning general rules from a large number of cases, which becomes the primary requirement. A representative example of this is facial recognition, such as face recognition technology. However, to create an artificial intelligence service, we cannot rely solely on rule-based or data-driven methods. All approaches must be appropriately integrated and used together. Thus, there are various ways to develop artificial intelligence.

However, these things did not come easily as soon as we started working on them. The basic concept of artificial intelligence began to emerge in the 1950s, and the idea of artificial neural networks also started to appear during that time. But back then, it was just an idea, and as people gradually began to concretize it, the perceptron emerged, and as development continued, people faced significant difficulties in developing any artificial intelligence based on neural networks. This effort continued, but from the late 1960s, people began to increasingly recognize these issues, leading to a growing sense of skepticism. As time went on, the prevailing thought became that artificial intelligence was not going to work well and that developing artificial intelligence was indeed a challenging task.

Until recently, this continued into the early 2000s. So we refer to this period as the so-called dark age of artificial intelligence. The people listening to this lecture, being young, may not fully understand, but when I was younger, I didn't think artificial intelligence would change society or become widespread like it is today. At that time, artificial intelligence was seen as somewhat of a foolish computer, rather simplistic. This was because it was referred to as the dark age of artificial intelligence, and there was a skeptical view that AI would not succeed. When we look back at the past of artificial intelligence, it was very difficult to solve complex intellectual problems, so we relied on simple rule-based systems.

As I mentioned, artificial intelligence can be categorized into rule-based and data-driven systems. At that time, there wasn't much data available, and the computers capable of processing it were not powerful enough. Therefore, most of what we could do was based on a few simple rule-based systems. Simple rules or search-based artificial intelligence were preferred, making it impossible to describe complex intelligence. For example, the artificial intelligence that emerged in the 1980s or 1990s was like a washing machine. As you know, an AI washing machine is not capable of performing complex tasks. It simply adjusts the amount of water and the spinning speed based on the laundry load and the washing method. And then there are the robo-advisors that are currently being talked about in the financial sector.

Now, there are articles coming out about the development of artificial intelligence, similar to its early forms, such as brain-like systems. However, in today's context, it is evident that these were very simplistic forms of artificial intelligence. Most of the artificial intelligence we have seen in the past was of this nature. While that was the case for artificial intelligence, the reality is that the very concept of informatization has rapidly advanced since the 1990s, particularly in what we call the IT sector. The 1990s marked the arrival of the informatization era. For example, computers have developed at an astonishing speed. This aligns with what is known as Moore's Law, which states that the number of transistors in a computer doubles approximately every two years. As a result, computers have improved significantly, leading to a dramatic increase in computing and processing speeds, along with advancements in data storage.

The capacity has also greatly increased, and at the same time, the prices of computers have significantly decreased. As the internet has developed, the amount of data that can be processed electronically has gradually increased. In the past, most data was not processable electronically and was simply on paper. However, nowadays, there is hardly any paper data; everything is stored on computers. Furthermore, the vertical integration and sharing of scattered data has become possible. In the past, transferring data required physically moving documents, but now it can be easily done with an email over the internet. Additionally, there have been advancements in data processing techniques. Methods for storing, managing, and analyzing large volumes of data have been developed, and new predictive techniques for data have been discovered. Through this process, arriving in the 2000s...

We have entered an era referred to as big data. You have probably heard a lot about big data by now. So, if we take a moment to look at what big data is, it essentially means that there is an enormous amount of data. By applying new analytical methodologies and management tools to this complex, large-scale data, we can uncover new information or insights that we previously could not find. This is the fundamental concept of big data. When we talk about big data, we often use the term 3Vs. So, what are the three Vs? The first is Volume, which refers to the increased scale of data. The second is Variety, indicating that in the past, data was primarily just numbers.

데이터의 형태가 대부분 정형 데이터였는데, 요즘의 빅데이터는 숫자 외에도 언어, 이미지, 비디오 등 우리가 소위 말하는 비정형 데이터가 많아졌다는 것입니다. 즉, 데이터의 형태가 다양해졌다는 이야기입니다. 비정형 데이터가 많아졌고, 그 다음에 '벨로시티'라는 표현을 쓰는데, 벨로시티는 약간 이해하기 어려울 수 있습니다. 데이터는 어떤 주기를 가지고 있습니다. 데이터를 처음 생성하고, 그 다음에 데이터가 유통되며, 사람들에게 소비되고, 다시 우리가 재저장하고 공유하는 이러한 과정을 거치는데, 이 과정의 주기가 점점 빨라진다는 것이 바로 벨로시티입니다. 그래서 요즘 들어 여러분들이 쉽게 알 수 있는 것은, 옛날에는 동영상 하나를 찍으려면 비디오 카메라로 먼저 찍은 다음, 그것을 컴퓨터로 옮기고, 다시 편집을 한 후 인터넷에 올리는 과정이 필요했습니다.

It took a very long time, but nowadays, you can take a picture with your phone and immediately upload it to the internet, allowing others to see it and respond right away. This is what we mean when we say that the cycle has accelerated. There is big data, which is typically represented by the 3Vs. So, what is important in big data? It is primarily about data collection: how to effectively collect large-scale data. Next is data management: how to manage, store, and then distribute this large-scale data to people. Finally, how to analyze this data and find new value that was not previously discovered in existing data. These issues have become the main concerns of big data. Big data started to emerge in the late 1990s, and as IT began to develop, it really took off in the 2000s.

Starting in the mid-2000s, the development of the so-called big data and the information age led to the emergence of deep learning. The person who actually spearheaded this deep learning revolution is Geoffrey Hinton from the University of Toronto, often referred to as the father of deep learning. He introduced the deep belief network in the mid-2000s, marking the beginning of the deep learning era. As these deep learning models were developed, they sparked a boom in artificial intelligence. Although I will discuss this in more detail during the lecture, deep learning models are characterized by the presence of multiple hidden layers that are not visible externally, forming what we call deep models.

We have now started to refer to it as a deep model. Usually, when we look at it, this is part of a neural network that you saw earlier. When you observe this neural network, you will see that some known information is input, and then the output yields the judgment or result we need. For example, if a photo is input, the output might conclude whether the subject is male or female. In this process, various neurons of the neural network extract information, ultimately providing us with something that enables our judgment. However, the values within these layers and the roles they play are not easily visible from the outside. We refer to these layers as hidden layers.

The term hidden layer is often used, and expressions like latent layer are also frequently mentioned. The existence of such layers is what we refer to as a deep learning model. In this deep learning model, hidden nodes or variables extract the necessary information for judgment through a learning process. There is a difference between traditional approaches and deep learning approaches. In traditional approaches, human experts select useful information and compute it from the data to make judgments. For example, before deep learning, to determine whether a photo depicted a man or a woman, a human expert would...

We can identify several distinguishing features. For example, the length of hair: typically, women have longer hair while men have shorter hair. So, let's base it on hair length. We will extract the hair length from the photo. After measuring the length of the hair, we can use that information to help determine whether the individual is male or female. However, deep learning models do not operate this way; instead, they internally search for latent variables that help in making judgments. We will gradually discuss these aspects further. 

When we look at examples of image recognition using deep learning, we can see that when a certain image data comes in, this image information goes through various layers, and we analyze the lines and features present on our faces.

As these lines change, they become parts of a face, such as the nose and eyes, and this ultimately forms the final facial appearance. This process goes through the layers I mentioned earlier. As the information passes through these layers, it gradually becomes more refined. Then, I can make a judgment about whether I know the person or not. With these advancements, we are now able to perform image classification tasks that were previously impossible. For example, determining whether an image depicts an airplane, a truck, or a frog was very challenging for computers in the past. However, with the advent of deep learning models, the performance of computers in these tasks has significantly improved.

Before deep learning, when we tried to solve these problems, the accuracy rates were around 25% to 28%. However, after the introduction of deep learning models, when we have a specific image, such as determining whether the object in the photo is a frog or a horse, the accuracy dropped significantly from 25% to 16.4%. This drastic drop occurred with the emergence of the AlexNet model in 2012, which was developed by Geoffrey Hinton's group. Following that, the accuracy continued to decrease to 11.7%, then 7.3%, 6.7%, and finally 3.5%. It is not an exaggeration to say that the problem of classifying images has nearly been solved, as the introduction of artificial intelligence deep learning models has shown clear effectiveness. As a result, through this process of artificial intelligence, starting from the mid-2000s, so-called...

Many people are starting to think that the spring of artificial intelligence is now upon us. With the development of deep learning models and artificial intelligence, we have made significant progress in areas such as speech recognition, computer vision, and natural language processing, which involves handling human language. Furthermore, artificial intelligence has expanded its role across various fields, including applied mathematics. 

So, the discussion revolves around the so-called spring of artificial intelligence, with some saying that summer will come soon. Some believe that artificial intelligence is already hot and that we are already in summer, while others think it hasn't fully arrived yet and that it will become more prevalent in our lives. There are diverse opinions: some think summer has already arrived, while others believe it will come soon. Regardless of differing views, it is clear that artificial intelligence is becoming the dominant trend.

The fact that it has been implemented means that it cannot be realized now. So, as we have been discussing, we have talked a lot about artificial intelligence, machine learning, big data, and deep learning. Now, I would like to explain how these concepts are related to each other. First, there is the broad framework of artificial intelligence. The methods to implement this artificial intelligence, as I mentioned earlier, can be broadly categorized into two types: rule-based and data-driven. The data-driven approach is what we refer to as machine learning. Therefore, machine learning is part of artificial intelligence. However, since machine learning is inherently data-driven, it requires data. If there is limited data, we can create a simple model to solve easy problems. However, if we want to solve more complex problems, we need more data.

복잡한 모델이 필요하고, 그러면 더 많은 데이터가 필요합니다. 그러니까 데이터가 굉장히 많이 필요한 경우도 있어요. 그러다 보니까 자연스럽게 우리가 소위 이야기하는 빅데이터와의 관련성이 생기게 됩니다. 그래서 빅데이터를 보면, 빅데이터는 또 인공지능과는 조금 독립적으로 나온 개념입니다. 빅데이터는 데이터 자체를 가지고 나왔던 거죠. 빅데이터라는 것이 크게 또 있습니다. 그런데 이 빅데이터에서 아까 중요한 요소가 데이터를 어떻게 우리가 생성하고, 수집하고, 저장하고, 관리할 것인가 이러한 문제가 있고요. 그럼 이거를 가지고 어떻게 활용할 것인가라는 문제가 또 중요한 문제가 있습니다. 그럼 활용하려고 보니까 사실 어떻게 활용해야 되는지 좀 애매한 거예요. 그렇죠? 빅데이터는 이제 데이터의 활용이란 문제가 있었고, 기계학습은 또 어떤 문제가 있냐 하면, 데이터가 많이 필요하다는 문제가 있었죠. 그러다 보니까 빅데이터와 기계학습이 서로 연결되는 부분이 있습니다.

A complex model is needed, and that requires more data. Therefore, there are cases where a large amount of data is necessary. Consequently, a natural connection arises with what we commonly refer to as big data. When we look at big data, it is a concept that emerged somewhat

There will be parts that both sides can share, and this is precisely the part we are interested in. This part aims to solve complex problems that we couldn't address before using very large data. The most important tool that can be used to achieve this is deep learning, which we believe plays that role. Up until now, at this point in time, everyone is trying to solve many problems that couldn't be tackled before using deep learning. I have discussed how we have come to utilize deep learning so far, but now let's briefly talk about how artificial intelligence is expected to develop in the future.

In my opinion, artificial intelligence can develop in two major directions. The first is that people want to create more general artificial intelligence. We refer to this as artificial general intelligence, or AGI. Currently, we are unable to achieve this. If we look at current artificial intelligence, we see that each program is designed to solve a specific problem. For example, AlphaGo is needed to solve the game of Go, while ChatGPT is required for chatting. These are different programs. However, a human can do many things; while some people may be better at conversation and others may excel at Go, a single person can do it all. Therefore, we can envision a single AI that can perform tasks like a human.

Isn't it possible to solve various problems? This is precisely what general artificial intelligence does. The technical aspect required to achieve this is called multi-modal learning. Basically, current artificial intelligence learns from only one type of data. For example, there are AIs that learn only from images and others that learn only from human language, and they are separate. So, what is multi-modal learning? It involves learning from images, sounds, languages, and texts, and combining these diverse types of data. For instance, a representative example is image captioning. What is image captioning? It is about looking at an image and generating text that describes that image. To do this, the AI needs to understand both the image and the text, meaning it must comprehend both. This is what we refer to as multi-modal.

So, if you look at video Q&A or the latest version of GPT-4, you'll see that it can recognize images to some extent. All of these things fall into this category. Of course, we still have a long way to go before we reach what we consider true multimodal or general artificial intelligence, but these attempts are one of the directions in which we are progressing. Another direction we can see is artificial superintelligence. The term super might sound a bit strange, but in English, it’s called Artificial Superintelligence. What this means is that current artificial intelligence, as we mentioned, requires data. Right? It needs data. What is data? For example, if there is a photo, it looks at the photo to determine whether the person is male or female. In order to do that, it must already have a judgment about whether the person is male or female.

The results of judgment must be consistent. So, with results indicating this photo is a man and this photo is a woman, artificial intelligence learns to produce similar outcomes. But then, one might wonder who decided that this photo is a man and that photo is a woman. The answer is that it was humans who made those decisions. Humans have already established that certain photos represent men and others represent women. Consequently, as artificial intelligence learns from this data, it can only mimic the judgments made by humans. This process of training models in such a manner is referred to as supervised learning. We provide a large number of results from human judgments and instruct the AI to make similar judgments. This is the typical way we train artificial intelligence, and as a result, AI inevitably ends up making judgments that are similar to those made by humans.

But if we want to make judgments that are superior to those made by humans and endow artificial intelligence with intelligence beyond that of humans, what should we do? This means that future AI could potentially discover new solutions that humans have not yet identified. If that happens, it would possess intelligence greater than that of humans, which we could refer to as superintelligence. So, what do we need to achieve this? Of course, various methods could be attempted, but the approach we are currently taking is reinforcement learning. We will discuss reinforcement learning later, but it is known as reinforcement learning. Reinforcement learning is not based on data created by humans; rather, it involves certain rules or fundamental principles that govern the world. We have what we call the basic principles.

This is based on the principle of understanding how social phenomena develop on their own. The most representative example of this is AlphaGo Zero. To briefly explain AlphaGo Zero, the program that competed against Lee Sedol in 2016 was AlphaGo Lee. If we look at how AlphaGo Lee was created, it is known that it learned from the results of actual matches played between humans, using over 100,000 game records. This means it referenced how humans play extensively. However, if we observe the game of Go itself, there are fundamental rules that govern it. Once you understand those rules, you grasp the principles of Go.

If you understand that principle well, I believe it is possible to play Go well just by using that principle without looking at how others play. This is the basis of AlphaGo Zero. So, what happens with AlphaGo is that we only teach the computer the basic rules of Go. It only learns how to play Go. Then, two AlphaGo Zero programs play against each other. At first, they won't be able to play well; they'll play poorly. But eventually, one of them will win. As they continue playing, when someone wins, they analyze why that player won. Then, they make slight updates. Eventually, the other side will win, and they will analyze that as well and update again. This process continues.

As they continue to engage with each other, they mutually develop and enhance their skills. This method of development is precisely what AlphaGo Zero embodies. Interestingly, AlphaGo Zero, for example, initially plays very poorly. However, as it continues to play, its skills begin to improve. After just 36 hours, even without any basic knowledge or guidance provided by humans, AlphaGo Zero achieves a performance level equivalent to that of AlphaGo Lee. As it progresses, after 72 hours, when AlphaGo Zero competes against AlphaGo Lee, it wins with a score of 100 to 0, demonstrating significantly better performance. After continuing for 40 days, it surpasses all existing Go programs, becoming the best-performing Go program in the world.

It becomes possible to develop artificial intelligence that is far superior to humans without any assistance from humans, simply by laying down the fundamental principles of the process, without making any judgments about how humans can help. This is precisely what many people are currently considering as a pathway to superintelligence. By utilizing this approach, we are challenging ourselves to solve many problems that humanity has not been able to resolve thus far. Recently, there has been rapid advancement in such artificial intelligence, and it is expected to continue to evolve in the future. I hope that through this process, everyone can study and explore these fundamental concepts together.

데이터 학습과 지능 1 1

데이터 학습과 지능 1 1