2022/03/18

⏱ 15mins 

entering

안녕하세요. 오늘은 캐글 이야기를 해보려고 합니다. 요즘 AI 분야에 관한 관심과 흥미가 나날이 커지고 있죠. 데이터 사이언스와 AI 관련 서적이 끊임없이 출간되고, 수많은 교육 플랫폼과 유튜브엔 관련 영상들이 앞다퉈 올라오고 있습니다. (업스테이지 김성훈 대표님(a.k.a 성 킴)의 <모두를 위한 딥러닝 강좌>도 AI 필수 입문 강좌인 거 아시죠?) 그만큼 전공자, 비전공자를 막론하고 AI 공부 열풍이 퍼지고 있는 듯합니다.

Upstage, a leading company in the field of AI, is also sparingly investing in nurturing AI talents who will take off on the global stage. We are producing the best talents in the AI field through the 'Global Residency (GR)' program.

📌 Purpose of Upstage Global Residency Program

WE RESEARCH AND DEVELOP AI TECHNOLOGIES WITH PRACTICAL LEADERS WHO HAVE CREATED THE WORLD'S BEST AI SERVICES. (IMPLEMENTATION AND SERVICE OF VARIOUS LATEST DEEP LEARNING MODELS)

TOGETHER WITH LEADER SEONG-HOON KIM, THE WORLD'S BEST KAGGLE GRAND MASTER, WE PARTICIPATE IN INTERNATIONAL AI COMPETITIONS SUCH AS KAGGLE AND CHALLENGE OURSELVES TO OVERCOME DIFFICULT TASKS.

WE CONDUCT WORLD-CLASS AI RESEARCH WITH PROFESSORS AND DOCTORAL STUDENTS AT HONG KONG UNIVERSITY OF SCIENCE AND TECHNOLOGY, KAIST, AND SEOUL NATIONAL UNIVERSITY.

This episode was written based on upstage talk by Upstage Yunsu Kim, who joined the GR program and became the 6th Grand Master in Korea as part of the upstage AI challenges team. After first encountering Python in a liberal arts class, Yunsu Kim, a talented person from the Department of Humanities who entered the field of artificial intelligence in earnest, captured the process and effort of becoming the Kaggle Grand Master.

In 2021, Kim Yun-soo's name was proudly ranked sixth on the list of Kaggle Grand Masters, which were only five in Korea. The Upstage AI challenges team also includes Kim Seong-hoon, Korea's third grand master. Last year, we all celebrated together the moment when a Kaggle Grand Master from the upstage was born, succeeding him.

WHAT WAS THE SECRET TO BEING ABLE TO BECOME THE KAGGLE GRAND MASTER, YOONSU, WHO WAS NOT AN AI MAJOR? THOSE WHO ARE NEW TO AI, AS WELL AS THOSE WHO ARE AIMING TO CHALLENGE THE KAGGLE COMPETITION, PLEASE PAY ATTENTION TO YUNSU'S STORY.

3 THINGS THAT HELPED ME WHEN I FIRST STARTED STUDYING MACHINE LEARNING AND AI

After being introduced to the liberal arts class, I started self-taught with interest in programming. As a non-major, when I first started studying AI, three things helped me a lot: English, Google, and Coursera . These three things have been a great organic help to my studies.

First, it is 'English' . Most of the AI technology, papers, and coding-related materials are in English, so I really felt the importance of language. Of course, since I was from a foreign language high school, I had a certain level of English, so the language burden was somewhat less. English is very helpful when studying in a specialized field, so if you build up your skills, it will definitely help you in your technical studies.

Next is 'Googling '. When I faced a problem I didn't know, I was able to find the answer more quickly through a Google search. Google was the best teacher. Searching in English can help you get more quality information and is more efficient because it gives you more options for answering code errors. In the field of artificial intelligence, the latest technology is pouring out every moment, so quick information update is essential. That is why there is no teacher as good as a search engine that provides a lot of information.

Last but not least is Coursera . Coursera, an online lecture platform, provides high-quality lectures in various fields. Among them, I took Python for Machine Learning, Deep Learning, and Data Science by Professor Andrew Ng. As many lectures are conducted in English, it seems that if you can listen in English, you will have more lecture options.

When is the Kaggle Contest Challenge?

For the first few months, I took Coursera classes and worked on personal projects. After that, I went straight to the Kaggle competition. It was about 9 months (January 2019) after I started listening to Professor Andrew Ng's lecture in April 2018. I quickly learned the theory through online lectures, and spent most of my time in Kaggle competitions to learn machine learning in practice. Everyone has a different learning style, so there is no right answer that everyone should be like this. However, as I spent a lot of time concentrating on the competition, it seems that I was able to improve my skills and competency more quickly through failures and successes. As Kaggle is a competition where real companies have to come up with solutions to the challenges facing them, it was also possible to focus on areas that could be solved with artificial intelligence.

Kaggle is divided into real competitions and tutorial competitions. Before participating in the full-scale Kaggle competition, I participated in three tutorial competitions to have time to adapt and practice. In the first competition, of course, the result was not satisfactory. However, as the frequency of participation in competitions increases, you gain a sense of the techniques commonly used in Kaggle. From that moment on, I was able to show a better performance. So in a year and a half, I participated in 8-10 competitions solo, winning 6 silver medals.

Another big turning point is coming regarding the performance of Kaggle competitions. I joined the Upstage AI challenges team in 2020. During my internship, I won my first gold medal in the Kaggle competition with my teammates, and I became a grandmaster in August 2021. By continuously improving our capabilities in the AI challenges team, we were able to achieve a record high ranking of 22nd (top 0.013%).

If you want to know more about Yunsu Kim's AI study process, check out [Upstage Talks] Until Undergraduate Liberal Arts Student Becomes 24th Kaggle Grand Master in the World !

Purpose of Participating in Kaggle Contest and What You Get

If there is one thing that has been gained through active participation in Kaggle competitions, it is 'growth as an AI developer' . As I said above, I was able to test my skills to my heart's content through actual practice.

Kaggle has several mechanisms to motivate participation. One of them is raising the leaderboard rankings and tiers . It's like an RPG game. Just as you invest a lot of time in the game to increase your ranking, Kaggle will need to continue investing in order to increase your rank. I was able to feel satisfaction and fun while experiencing the experience of raising the ranking so that I could work hard. This little fun naturally led to interest and passion for the latest technology. Contemplating different approaches to solving problems, more trials and efforts were made. After that, you will feel that your skills are gradually filling up, and it is a fun structure where you can show your value as a developer by taking a high ranking in the Kaggle contest where you can prove your skills externally, so many people seem to be actively participating.

There are other benefits to participating in Kaggle competitions. Discussions, a place where great developers all over the world share their code and ideas, can give you tons of insights and new programming ideas . Developers share their questions and difficulties while solving problems, and other developers will post their own alternatives and answers. In it, you can find different problem approaches and definitions to broaden your perspective on problems. This is a very helpful part for the growth of novice developers. Also, since there are various problem domains presented in the contest, if you participate in various contests (Vision, NLP, Tabular, etc.), you can learn SOTA skills for each field. As such, for developers, Kaggle competitions seem to function as a place of learning and growth.

📌 What is Kaggle Grand Master?

To become a Grandmaster, you must earn 5 gold medals. It must contain at least one Solo Medal. (Kaggle can participate solo or as a team) The total number of participants in the Kaggle competition is about 170,000. Of these, there are only 248 grandmasters worldwide. That's a great title that only a few people can get. The Kaggle competition is divided into a total of 5 grades (tiers). Novice → Contributor → Expert → Master → Grand Master (highest grade)

The driving force behind becoming a Kaggle Grad Master

There are four main factors that have made me a Kaggle Grand Master. Passion, time, equipment, great team .

First, it is ' passion' that comes from fun. As I said before, I was able to participate in a fun way because there is a ranking system that returns just as much as my efforts. Of course, it is not easy to achieve good results in a short period of time. Even if you experience many failures, I believe that good results can be achieved in the end if you have the motivation and passion to participate consistently.

Second, 'time' . In order to focus on one Kaggle competition, you need time and space to devote a month or two to Kaggle alone. There is no standard on how much time you should invest each day. In my case, I invested most of my daily life in Kaggle until I became a Grandmaster. Of course, it seems important to find a balance by setting your own standards in parallel with other work.

The third is ' equipment' . Equipment is important when participating in image competitions or competitions that deal with large data. It can be difficult to reach the top without equipment that can perform well enough. For example, it takes one day to complete one experiment with one GeForce RTX 2080 Ti. On the other hand, if you have 4 RTX 3090s, you can finish it in 4 hours. When a large amount of GPU memory is required, it is sometimes difficult to properly learn. However, it does not seem necessary to have all the equipment when starting out. I started with a single 2080 Ti, AMD Ryzen CPU, and 64GB of RAM. If equipment is hard to come by, Google Colab TPUs may be an option. Joining the upstage GR team has helped me a lot to achieve a more satisfactory performance with good equipment.

The last one is ‘ a great team member ’. You can participate in the competition solo or with a team. Each has its pros and cons. Participating solo has the advantage of greatly improving your skills because you have to solve all problems by yourself, while participating in a team has the advantage of being able to come up with various ideas and attempts with your team members. Disadvantages may be the opposite of advantages. Since there are pros and cons like this, I think it is important to try different challenges depending on your personality. I like both of them, and we were able to compete as a team and win our first gold medal. In the Upstage AI challenges team, I was able to learn a lot through collaboration with team leader Kim Seong-hoon (aka Lime Robot_World's 12th highest ranking, Korea's third Kaggle Grand Master) and colleagues. This experience helped me a lot in improving my machine learning knowledge and capabilities as well as achieving good results in Kaggle competitions.

concluding

So far, we have told you the story of Kaggle Grand Master Yunsoo Kim. We learned about the journey from first encountering AI to joining the Upstage AI challenges team and becoming a Kaggle Grand Master. Upstage is always supporting Yunsu, who, as a non-major, has put more sincere effort into studying AI than anyone else. If you are curious about more of Yunsu's story, please check out the Upstage Talk, which contains Yunsu's story and various study tips.