About Me :

I am interested in designing "reliable" machine learning methods that make confidence to deploy them in a real world.
         - Interpretability : reasoning the model's behaviour gives trust to human experts.
         - Uncertainty : confidence of the model's prediction provides human whether to believe the forecast.
Therefore, my area of research is Explainable AI, uncertainty, and calibration.

Bio :

- Ph.D. student in CS, University of Tuebingen and INRIA with ELLIS program (Aug. 2021 – Present)
         - Advisor : Prof. Zeynep Akata and Prof. Cordelia Schmid

- Master’s degree in EE, Seoul National University (2018 – 2020)
         - Advisor : Prof. Jungwoo Lee
         - Best dissertation award

- Bachelor’s degree in EE, Seoul National University (2011 – 2018)
         - Full tuition, National Scholarship for Science and Engineering
         - 2-year absence to fulfill military duty (2013 - 2015)

Work Experience :

- Research intern at Alsemy (Mar. 2021 - Jun. 2021)
         - Startup company developing AI-based semiconductor modeling software

- Research intern at NAVER AI Lab (Aug. 2020 - Feb. 2021)
         - Mentor : Dr. Junsuk Choe and Dr. Sungjoon Oh


Variational Saliency Maps for Interpretability

TL;DR: Uncertainty for attribution that explains the model's prediction.

Jae Myung Kim, Eunji Kim, Seok Hyeon Ha, Sungroh Yoon, and Jungwoo Lee


Distribution-Driven Disjoint Uncertainty Estimation for Regression

TL;DR: Well-calibrated regression uncertainty estimation.

Jaehak Cho, Jae Myung Kim, and Jungwoo Lee


Keep CALM and Improve Visual Feature Attribution

TL;DR: Self-explainable model by simple modification of CAM, but with better explainability.

Jae Myung Kim*, Junsuk Choe*, Zeynep Akata, and Seong Joon Oh
[Arxiv] [code]


REST: Performance Improvement of a Black Box Model via RL-based Spatial Transformation

TL;DR: Studying the robustness of a black-box model to geometric transformations.

Jae Myung Kim*, Hyungjin Kim*, Chanwoo Park*, and Jungwoo Lee
[AAAI 20]


Exploring linearity of deep neural network trained QSM: QSMnet+

TL;DR: Better estimation of QSM, a quantitative approach for measuring magnetic susceptibility using MRI.

Woojin Jung, Jaeyeon Yoon, Sooyeon Ji, Joon Yul Choi, Jae Myung Kim, Yoonho Nam, Eung Yeop Kim, and Jongho Lee
[NeuroImage 20]


Sampling-based Bayesian Inference with Gradient Uncertainty

TL;DR: Efficiently predicting a predictive uncertainty by incorporating the concept of gradients uncertainty into posterior sampling

Chanwoo Park, Jae Myung Kim, Seok Hyeon Ha, and Jungwoo Lee
[NeurIPS 18 Workshop]