Currently, I'm interested in deep learning, especially representation learning in computer vision, and real-world applications (e.g. unmanned vehicle, autonomous driving).
Feel free to check out my
CV
and drop me an
e-mail
if you want to chat with me!
Research Student | Electronics and Telecommunications Research Institute
Jan '21 — Feb '21
I worked at Defense & Safety ICT Research Department of Intelligent Convergence Research Laboratory.
We built a deep learning model that predicts crime types through police reports.
Research Intern | Ewha Womans University
June '20 — Aug '20
Working under the supervision of Prof.Hyunseok Park at the Bioinformatics Laboratory.
Finding insights from bioinformatics journals by using NLP methods.
(*) indicates equal contribution.
STAR-SUM: Structure-aware Video Summarization Jinhwan Sul*, Yonsoo Kim*, Donghwa Kim, Joonseok Lee Submitted to CVPR, 2023
An Empirical Study on the Korean Text Classification based on Active Learning with KorBERT Wonjoo Park, Yonsoo Kim, Myungsun Baek, Yongtae Lee. The 2nd Korea Artificial Intelligence Conference, 2021
Organizing an in-class hackathon to correct PDF-to-text conversion errors of Genomics & Informatics 1.0 Sunho Kim, Hyunseok Park, [and 43 others, including Yonsoo Kim] Genomics & Informatics, 2020
Active Learning Apparatus, Apparatus and Method for Sampling Data used for Active Learning Yonsoo Kim, Wonjoo Park, Yongtae Lee Apr, 2021. Patent pending
Dean’s List, Ewha Womans University, 2018 — 2021
2nd Place in SW Startup Competition, Ewha Womans University, 2021
4th Place in Best Poster Awards of Graduation projects, Ewha Womans University, 2021
3rd Place in Finance Data Contest, Financial Security Institute, 2021, (among 400+ teams)
Finalist in Animal Datathon Korea, Animal Tech Korea, 2021
LEAP Student Club Grants, Ewha Womans University, 2021
Honorable Mention in Industrial Control System Threat Detection AI Competition, National Intelligence
Service(NIS), 2020, (top 10 among 260+ teams)
STAR-SUM: Structure-aware Video Summarization
CVPR 2023 (Under submission) Dates of Participation: Mar '22 — Present
In this paper, we proposed a structure-aware transformer, which hierarchically processes each semantic level composed by machine-generated semantic units.
We also reset evaluation settings for fair comparison and suggested experiments related to video summarization subjectivity, and its positive results confirmed the model possessing generalizability.
Token Sampling for ViT
Dates of Participation: Sep '21 — Mar '22
We proposed a token sampling module for vision transformers to focus on core tokens rather than giving all tokens the same weight.
We built a trainable kernel function that computes probabilities of tokens in the image.
By doing this, the model could select the most important tokens of regions. This module can be attached to any vision transformer models, such as Vanilla ViT, Swin Transformers, TimeSFormer, MViT and etc.
Data Analysis Tool using Machine-Generated Visualization
Dates of Participation: Mar '21 — Dec '21
In this work, we developed a data analysis tool that recommends data plots generated by a deep learning model(Data2Vis), thereby providing data analysis reports in dashboard form.
The project recognized as a high-quality UI/UX for broad accessibility from a user survey and resulted in 2nd place in the SW Startup Competition and 4th place in the Best Poster Awards of Graduation projects.
One Line A Day
Dates of Participation: July '21 — Dec '21
We created a mobile application with a function that summarizes in one line through a text summarization model after a user writes a diary. My role was a machine learning engineer and backend developer. I implemented abstractive summarization API using Pororo, a Korean text summarization model.
Active Learning on Crime Classification
The 2nd Korea Artificial Intelligence Conference, 2021 Dates of Participation: Jan '21 — Feb '21
In this work, we conducted a crime classification project based on crime reporting messages using KorBERT.
I suggested a new active learning algorithm where the model is proportionally fed to learn high confidence data as well as low ones.
The performance was improved macro F1 11.6% higher and weighted F1 4.9% higher for the same training time.
We published a paper in Korea AI conference, got a patent pending, and further exhibited by KOREA AI EXPO 2021.
Sign Language Education Web Application
Dates of Participation: Sep '19 — Aug '20
We developed a sign language “education” web that could lower the entry barrier of sign-language for the public.
My team designed it to recognize the user’s hand gestures using a real-time object detection model through a webcam, which would make the user practice each alphabet of sign-language ten times.
For this, we trained a YOLO model from scratch, using our labeled data from over 20,000 images.
This template is a modification to Rishab Khincha's website.