• 한국어
    • 日本語
    • 中文-繁體

    2023.01.30 The Originality

    AI Researcher, Anomaly Detection, Hoyeop Lee

    <THE ORIGINALITY> is a series about NC’s new generation — they are immersed in their jobs where they find inspiration. They pave the way towards excellence and then aim even higher.

    People at NC freely express themselves and achieve growth by challenging themselves to new experiences.

    When faced with a problem that is difficult to solve, think deeply. If it takes too long, it means you don't know the field well. Think again after researching related studies or its background knowledge to the point where you feel comfortable about the problem. Then write down the answer. For me, this was the fastest way to improve myself.

    AI Researcher, Anomaly Detection, Hoyeop Lee

    Anomaly Detection, Applied AI Lab

    Applied AI Lab is an organization at NC that researches AI technologies which can be applied to new businesses in various fields, such as, media and finance. The Anomaly Detection Team, affiliated with the Applied AI Lab, researches AI with the goal of detecting anomalies in various patterns of data and providing meaningful information to users based on it.

    One of the methodologies that has recently been attracting attention in the academic world is Graph Neural Network (GNN) technology. In order for AI to detect anomalies, it must first learn, and GNN is an effective algorithm for learning graph-structured data. A graph consists of nodes and edges connecting those nodes. It is easy to understand if you equate them to friendships on social media or subway maps. Graphs are a common data type, similar to images or voices, in everyday life, and easily represent complex relationships. Currently, our team is defining anomalies based on GNN and conducting research.

    The Two Opposites of Normal, “Abnormal” and “Not Normal”

    There are two main methods to detect anomalies. One is to distinguish between “abnormal” and “normal,” and the other is to distinguish between “not normal” and “normal.” They may sound the same, but if the former has two classes, “abnormal” and “normal,” the latter uses only one class.

    Let us take a cup for example. The way to determine if it is “abnormal” or normal” is to learn that an intact cup is recognized as normal and a dented cup is abnormal. On the other hand, the method of classifying “normal” and “not normal,” is by teaching AI that anything but an intact cup is not normal, so that later it can detect the “not normal” state.

    The disadvantage of the “abnormal” and “normal” method of classification is that the computer will only learn that dented cups are abnormal and will not recognize stained cups or cups with two handles as abnormal. Therefore, current research focuses on distinguishing normal and not normal.

    A Reminder to Prepare for the Abnormal

    More often than not, there are more cases when people tend to ignore abnormal behavior. Especially, the significant abnormalities that would greatly affect our lives require sufficient preparation. However, if we fail to detect them, we will be vulnerable.

    Our team strives to thoroughly inform people of parts they may have unknowingly ignored, and to help them fully prepare. The value that Anomaly Detection pursues lies there.

    Viewing Relationships through the Eyes of Technology

    How to Explain a “Cup” to a Computer

    My older brother is fascinated by physics. Influenced by him, I became interested in discovering connections between things that were irrelevant to each other or viewing macroscopic phenomena by analyzing microscopic relationships.

    In a nutshell, AI technology is the process of making artificial intelligence understand the myriad complex relationships that exist in reality. People naturally recognize a cup as a cup when they see one. However, it is a very difficult process when it comes to computers. When we input information that helps a computer recognize a cup as a cup, the learning process must be modeled until the output is “this is a cup.” It was fun to play with questions like “Why do computers have such a hard time when it is simple for humans?” and “How can we explain that a dented cup is not normal?” That's how I walked the path of an AI researcher.

    Creating Something Out of Nothing through Interaction

    The essence of research is to solve problems for which the answer is unknown. Then comes the key element, “interaction.” The power of interaction comes from deducing new ideas derived from sharing information that is already well known. To add to that point, I think NC is a place with the best environment for researchers. The company culture is horizontal and flexible, and there are many outstanding researchers who can exchange and develop each other's ideas. Productive interactions flourish here every day.

    As a team leader, I also try to build an environment where team members can grow. You know what they say about knowledge being fundamental to all achievements and progress. I continue to stay up to date with the latest papers, especially on machine learning. Also, I often spend time sharing ideas with my team members. I give feedback on results and provide opportunities to create synergy by encouraging my team members to have frequent coffee breaks together even if I’m not present.

    Successful Research and the Sense of Accomplishment

    It felt like every work I did in graduate school was up in the air because it all concluded with my thesis. I was curious how the technology I researched could be applied to various services in real life.

    Before joining the current team, I was researching an algorithm that clustered news, and an algorithm that identified users' preferences and offered recommendations. At the time, I wrote a paper on recommendation algorithms and filed in for a patent. After a while, another organization contacted me asking permission to use my findings for their service. I felt like I took one step forward as a researcher when I realized that the technology I developed could be useful to someone else.

    I am still working with the organization that contacted me and continuing the project. I also expect Anomaly Detection to expand into a technology that delivers meaningful information to a lot of people.

    Feynman's Secret, “Think Very Hard”

    Do What Is Fun, Professional Growth Will Follow

    I tend to pursue entertainment in everything. In college, I was an inquisitive student. Usually, people often passed by without giving the school bulletin board a look, but I always took interest in what they were advertising. I participated in contests that intrigued me, or applied for research on interesting topics.

    It was the same with NC. Given the nature of a company, it is not possible to do only fun things. Even so, I tried my best to find some excitement in the work I was given, as I was certain that I would learn and gain something from every task I had to do. Excitement and curiosity were the driving force behind my growth.

    Repeat Until You Solve the Problem

    American physicist Richard Feynman, who won the Nobel Prize, said that he went through the following three steps1 to solve problems.

    When solving a problem, some people tend to try out different approaches and analyze the results later — it is a process of finding the answer through countless trial and error attempts. However, in Feynman's algorithm, problems need to be thought through before getting to problem solving which is different from taking a trial-and-error approach to the issue.

    If it takes too long to come up with an answer, it probably means that you don’t know the field well. In that case, you must research related studies, build background knowledge, and think about the problem. For me, coming up with my own answer through this process was the fastest way to improve as a researcher.

    1. 1. The Feynman Problem-Solving Algorithm

    1) Write down the problem.
    2) Think very hard.
    3) Write down the solution.

    Reasons Produce Reliability

    Next year, I want to develop a model that will be able not only to detect an anomaly but also explain why it is considered an anomaly. Rather than simply conveying information, people will trust the information more when I can explain what algorithm this information was derived from and why it is classified as an anomaly. I want to become a researcher who can explain the process of detecting anomalies in a language that even non-specialists can understand.

    * The content stated in this interview is the personal opinion of the interviewee and does not represent the official position of NCSOFT.