Daejin Kim

I am an M.S student at DAVIAN Lab (Advisor: Jaegul Choo), part of the KAIST AI at Korea Advanced Institute of Science and Technology.

I am currently working on computer vision and time series forecasting. Most recently, I have been interested in finding ways to guarantee the faithfulness of the visual explanation in deep-learning-based models.

Email: kiddj@kaist.ac.kr

[CV] /  [Portfolio (Korean)] /  [LinkedIn] /  [GitHub]

Publications

(C: peer-reviewed conference, J: peer-reviewed journal, * = equal contributions)

[C4] Mining Multi-Label Samples from Single Positive Labels

· Propose a novel way to draw samples of joint classes (e.g., 𝐴 ∩ 𝐵) using only single positive labels (e.g., 𝐴, 𝐵).
· Estimate the conditional density of (non-)overlapping classes using MCMC method with logits of classifiers.

[C3] WaveBound: Dynamic Error Bounds for Stable Time Series Forecasting

· Introduce the dynamic error bounds to address the overfitting issue in time series forecasting.
· Propose a novel regularization method that estimates the training loss inevitably occurs in noisy patterns.

[C2] Residual Correction in Real-Time Traffic Forecasting

· Identify that recent deep-learning-based traffic forecasting methods does not handle the residual autocorrelation.
· Propose a simple add-on module to reduce residual autocorrelation and consistently improve the performance.

[C1] Not just Compete, but Collaborate: Local Image-to-Image Translation via Cooperative Mask Prediction

· Improve the existing face editing methods by preserving the attribute-irrelevant regions using Grad-CAM.
· Propose a novel loss that allows the generator and the discriminator to collaborate.

Unpublished Work / Projects

Your Lottery Ticket is Damaged: Towards All-Alive Pruning for Extremely Sparse Networks

· Explicitly handle the useless weights occurred by existing saliency-based pruning methods.
· Improve the performance of existing saliency-based pruning methods (e.g., MP, SNIP, LAP) at high sparsity.

Knowledge Distillation in BERT

· Implement a smaller version of BERT (with 4 layers and 8 attention heads) by using knowledge distillation.
· Significantly reduce the number of parameters of BERT while minimizing performance degradation.

Editable Text-Adaptive GAN

· Inspired by Text-Adaptive GAN and Editable GAN, propose a single GAN that can generate and manipulate images simultaneously for the given text prompt.

Experience

NAVER WEBTOON Corp.

DAVIAN, KAIST AI (Advisor: Jaegul Choo)

DIALLab, Sungkyunkwan University (Advisor: Jongwuk Lee)

Purdue University

Participate in the IITP Purdue Capstone program at Purdue College of Information Technology, sponsored by the Korean IITP (Institute of Information & Communications Technology Planning & Evaluation).

Hanbom High School

Awards & Certifications

Dean's List

First place at KAIST AI World Cup AI Commentator Session

Design and source code from here