Program Operation Performance

  • KAIST StartupTing X Upstage Review
  • 창업원
  • 2022-10-30 22:01:54
  • 362

https://youtu.be/nq6mO013038

Today’s StartupTing is joined by Upstage, an AI startup created by AI core developers from Naver, Kakao, Nvidia, and Google. CEO Sung Kim is known for his deep learning lecture, and is currently a professor at the Hong Kong University of Science and Technology(HKUST). If you want to learn about AI and receive mentoring, stay focus on Upstage’s StartupTing.



Career Building in the AI era – CEO Sung Kim

CEO Kim started the StartupTing with a lecture on career building. The lecture will be described from a first person perspective below.


The domestic industry has changed from physical, digital to AI Beneficial. I thought about how to AI transform the myriad of problems in the industry, and I felt the need to create value for a new field of AI modeling, engineering -> cost reduction. I felt that AI planners were necessary to create this value, so I started Upstage. I wanted to create a self-improving AI (SAI) platform, so I named the company Upstage.


We create AI models from a base model that recognizes data, and are planning to make a platform as well. Upstage is a startup founded by AI developers and business leaders with experience in AI services. We also look for talented AI members that will join us in our AI platform.



A story never told of school, company, and startup – CTO Hwalsuk Lee

Before CTO Lee started on his AI lecture, he told his career story as a KAIST senior.


[Career story of CTO Hwalsuk Lee]


After obtaining a Ph.D. on video codec at KAIST, I worked at Samsung Techwin for 15 years. While working on CCTV technology, I read a Deep Mind paper and felt the urge to pursue AI, so I bought GPU and started research. To focus on AI, I moved to NCsoft to develop a bot for one-to-one battles, researched generative models for game development efficiency, and created lectures on autoencoders. In particular, the git repository consisting of generative models using Tensorflow, attracted attention, leading to a collaboration with Google and Reddit.


In order to provide AI services to more customers, I moved to Naver and advanced the OCR (Clova OCR) field. Due to this, I wrote 11 papers and also won the OCR World Championships. After being recognized the contributions and developing dozens of domestic and foreign services, I became responsible for the overall computer vision, including OCR, and joined CEO Kim to start a business.


[School vs Company: A perspective from AI development]


In school or research, one usually looks for a better model for a fixed dataset or evaluation. Research focuses on developing AI models that draws new technologies accordingly with a fixed number of test sets, test sets, and evaluation methods. However, in AI development, there is no fixed train set, test set, and evaluation method. Thus, there may be a gap between a quantitative assessment in the development environment (offline test) and in the real world (online test). Since quality of the service is important, it is critical to design offline tests similar to online tests, and derive specifications of the model.



[Company – AI Composition of organization]


You need members to prepare and manage the quality of the aforementioned data. The company needs a workforce that develops tools for efficient data and modeling, and people who manage the overall quality of models, like data curators, modelers, and IDE developers. If the tech team also serves AI models, more workers are required because there is more work to serve the model to end devices.


One must ensure that data accumulates and consider service requirements when developing services. Thus, the first thing to do when developing an AI model to a service is preparing training data sets. For example, one must define the type, quantity, and target data.


e.g. AI development that predicts Latex expressions when taking pictures of formulas that are difficult to input


To prepare a training dataset, an AI model must be designed, and to design an AI model, a training dataset is needed. It is important to converge as we repeat this process, and now there are a lot of data manufacturing companies that we outsource data production.


[Company vs Startup]


Startups go through the process of Scope Project-> Collect Data-> Train Model->Deploy in Production.


Startups nowadays have big data like any other companies. In the Collect data part, we have to see if data accumulates and check the quality of the data.


In the Train Model part, there are tools that efficiently use GPU, government funding, and startup support businesses that startups can utilize. However, the most important part is the Scope Project part. It is important to develop an AI model according to the service goal. The content below is a worthy consideration when selecting a company depending on the two types of businesses.


Business Centric AI: companies that have grown already – the business before AI was incorporated, existing modules, business flow do not change much


AI Centric Business: startups – place AI at the center and define a new flow (AI Transformation).


e.g. traditional car manufacturers vs Tesla -> or completely new services/businesses.


To summarize, companies like large corporations already have a set rule on AI, so it is difficult to change the stage considering the service stability. On the other hand, startups have no limit, so it is relatively free to create a new stage to introduce new technologies.



AI research and development at Upstage – AI researcher Sungjoon Park

AI researcher Sungjoon Park told in-depth stories about AI research and development at Upstage, focusing on how to process data and NLP. Referring to the StartupTing video will be useful to learn more about AI.


In 2018 Glue, in 2020 BART, and this year GEM appeared. Our team participated in the KLUE project (Korean Language Understanding Evaluation), with 31 co-researchers, 11 collaborators, and 10 sponsors.


AI developers should also consider ethical issues on data sets. It is important to think about how to deal with these issues from a model evaluation perspective. Recently, there are many competitions to elect AI developers and more AI experts are being trained. There was also a test on domain adaptation by Riiid.



Q&A and Discussion

CEO Sung Kim, CTO Hwalsuk Lee, and researcher Sungjoon Park each answered pre-registered questions based on the categories of AI career, AI development, and AI technology.


Looking over a few questions, to the question “Which is better – a company with excellent technology or a company that analyzes data well?, they answered that “a technology company can learn from people around them, and a data analytics company can try many things with the abundance of data, but to choose between them, we prefer a company with excellent technology”. To the question asking what non-AI majors should study to enter the AI industry, they answered that it is important to build a strong foundation with basic computer science courses. To the question asking what ethnical guidelines exist and which ethical value they consider most important, they answered that there are a lot of guidelines on research in universities and companies to refer to.


Besides pre-registered questions, there were many more questions asked live. In the StartupTing of Upstage, there were about 200 participants. There was even a reporter interested in startups that joined to write an article. The interest in AI is increasing as the AI era accelerates. We hope more AI experts appear in KAIST that will lead to even more startups.


Please look forward to and stay tuned for the next StartupTing!


첨부파일