Jaerin Lee

Varying optimization algorithms and learning rates on a deep learning task reveals a distinction between the problem-intrinsic trajectory and optimizer-specific edge of stability.

A greedy alignment principle turns manual optimizer hyperparameter tuning into a solvable filter–statistics alignment maximization problem.

The previous definition of the edge of stability mispredicts the stable sharpness with an optimizer-dependent systematic error; this work presents an interpretable alternative.

Faster grokking with a two-timescale momentum optimizer acting as a low-pass amplification filter for gradients.

Structure-aligned optimization schedule adds 1.4 dB PSNR in x2.5 shorter training time for image vectorization.

A 3D Gaussian splatting scene generation pipeline that recursively generates scenes from off-the-shelf diffusion models and a point cloud alignment algorithm.

By combining previously incompatible region-based control & fast schedulers, Semantic Draw turns any image diffusion model into a live paintbrush tool that let you brush semantic intent.

An alternative to RGB colors, aRGB representation equips pairwise per-pixel loss functions with structural information for better supervision of various image restoration models.

CIHR improves the deblurring quality of images by exploiting the inverse task of deblurring, namely, reblurring.

Highlighting that GAN discriminators are also data models, SRD is a discriminator exploiting three structural features of natural images: translation equivariance, rotation invariance, and hierarchy of scale.

An AIM challenge on example-based Video Temporal Super-Resolution for enhancing videos for better temporal resolution, held jointly with the 2nd AIM: Advances in Image Manipulation workshop, in conjunction with ECCV 2020 in Glasgow, UK.

Thesis advisor: Professor Kyoung Mu Lee. This work explores the optimal architecture for an image classifier by progressively growing a CNN with respect to intermediate activations, achieving better results than randomly wired networks.

An 18-month term of compulsory military service at the ROK Army AI R&D Center, completed with the final rank of Sergeant. The work spanned three projects commissioned by the Superintendent of the KMA (LTG ★ ★ ★), from proposals to deployments, with coverage in two documentary TV episodes from a state broadcaster, recognition through two Commander’s Commendation Ribbons (LTG ★ ★ ★, MG ★ ★), an award at the 5th Army Startup Competition (MG ★ ★), and one invited talk for the faculty members of the KMA.

A commendation from the Superintendent of the KMA, MG Jeon, Sungdae (★ ★), recognizing contributions to intelligent military surveillance through the project 'Development of AI Autonomous Surveillance System in Korea Military Academy Ammunition Storage Area.'

An unstaffed, fully autonomous surveillance system for military operations in the ammunition storage area of the KMA. The first-ever field deployment of deep learning algorithms in the ROK Army, commissioned by LTG Chang-gu Kang (★ ★ ★), the Superintendent of Korea Military Academy, and recognized with a Commendation Ribbon (MG ★ ★).

A real-time video restoration system for military surveillance cameras exposed to inclement weather. The first-ever military application of deep learning-based low-level vision algorithms in the ROK Army, commissioned by LTG Chang-gu Kang (★ ★ ★), the Superintendent of Korea Military Academy.

Title of the talk: 'High Performance Computing for Artificial Intelligence.' A one-and-a-half-hour lecture for the faculty members of the KMA at the 1st Korea Military Academy Artificial Intelligence Seminar held by the ROK Army AI R&D Center.

Episode title: 'Dreaming of the Army with Advanced Science and Technology, Military Science and Technology Researcher in Korea Military Academy, Part 2, I Am a Republic of Korea Soldier, episode 176.' Another half-hour documentary TV show about enlisted researchers supporting cadet education with teaching assistance and technology demonstrations on AI in the KMA.

Episode title: 'Dreaming of the Army with Advanced Science and Technology, Military Science and Technology Researcher in Korea Military Academy, Part 1, I Am a Republic of Korea Soldier, episode 175.' A half-hour documentary TV show about enlisted researchers in the Army and the first AI field application in the ROK Army.

A commendation from the Superintendent of the KMA, LTG Kim, Jeong Soo (★ ★ ★), recognizing contributions to intelligent military surveillance through the project 'Development of AI Autonomous Surveillance System in Korea Military Academy Ammunition Storage Area.'

An automated surveillance system integrating real-time multi-object tracking, super-resolution, and re-identification systems with a user-friendly interface. The first ever military application of AI in the ROK Army, commissioned by LTG Kim, Jeong Soo (★ ★ ★), the Superintendent of Korea Military Academy. The results were presented at the 21-2 semi-annual meeting of the Steering Committee of the KMA and recognized with a Commendation Ribbon (LTG ★ ★ ★).

Proposal title: 'Military Consultation Chatbot for the Army Mental Health Services Using Large Language Models.' Awarded by the Commanding General of the ROK Army Personnel Command, MG Jeon, Byung-hee (★ ★).