Ki-Ung Song
AI grounded in mathematics
Edge AI, AI for Science, AI for Finance
Deepest
Deep Learning Research Society at Seoul National University
Overview
Deepest gave me a way to keep doing serious deep learning work outside my job, through seminars, study groups, projects, and challenge teams with strong peers.
That mattered because it let me keep reading difficult papers, presenting technical topics, building systems, and testing ideas across a wide range of deep learning areas without letting my perspective collapse around only one line of work.
How I Engaged With Deepest
Through Deepest, I stayed involved in several different forms of technical work: weekly seminars, focused study groups, seasonal team projects, and challenge-based experimentation. The value was in moving repeatedly between reading, presenting, implementing, and competing rather than staying in only one mode of learning.
Selected Explorations
Seminar Presentations
Ongoing
A recurring part of Deepest for me was preparing and giving seminars on more mathematical or less commonly discussed topics in deep learning, including Neural Tangent Kernels, singular learning theory, SO(3)-equivariance, and deep learning for simulation.
Parallel Computing Study
Oct 2023 - Feb 2024
A study group focused on GPU computing, CUDA, and performance-oriented deep learning workloads, which helped strengthen my systems intuition around efficient deep learning.
NTIRE 2022 Challenge Team
Nov 2021 - Apr 2022
A competition-focused team project on image super-resolution, where we explored high-frequency detail and achieved strong workshop rankings. This work later led to a publication.
Offline RL Project
Nov 2021 - Feb 2022
An implementation-focused project for understanding offline reinforcement learning through direct experimentation in PyTorch, which made the topic much more concrete than reading alone.
What It Kept Alive
Deepest gave me a place to keep technical curiosity and ambition active outside my core employment. It helped me stay current, test ideas with strong peers, and maintain a wider sense of the field than day-to-day specialization usually allows.