Hanseul Cho

Hanseul Cho

PhD Student at OptiML Lab in KAIST AI. Working on ML/DL/Opt+Theory and LMs.

Hanseul Cho (조한슬) 👋

I am a Ph.D. student in the Optimization & Machine Learning (OptiML) Laboratory, where I am fortunate to be advised by Prof. Chulhee Yun at Kim Jaechul Graduate School of AI in Korea Advanced Institute of Science and Technology (KAIST AI). Previously, I completed my M.Sc. (in AI) and B.Sc. (in Math, minor in CS, Summa Cum Laude) at KAIST.

I am interested in a broad range of fields in optimization, machine learning (ML), and deep learning (DL), especially focusing on both:

  • Mathematical/Theoretical Analysis, and
  • Empirical Improvements (usually based on theoretical understanding).

Specifically, my research interests lie in the following subjects:

  • Understanding and mitigating the fundamental limitations of modern (large) language models (e.g., length generalization and compositional generalization of Transformer architectures)
  • Online learning & optimization under circumstance shifts (e.g., continual learning, maintaining plasticity of neural networks, reinforcement learning, streaming PCA)
  • Structured/constrained/multi‑level optimization (e.g., minimax optimization, bi‑level optimization, fairness in ML)

📰 Publications 📰

Please click the “Publications” tab above to look up the full list of my publications.
You can also find my articles on my Google Scholar profile.

‼️News‼️

  • 🗞️ [Nov. '24] Our paper on theoretical analysis of continual learning is accepted to JKAIA 2024 and won the Best Paper Award! 🎉 (See Publications for more details)
  • 🗞️ [Nov. '24] I'm selected as one of the Top Reviewers (top 8.6%: 1,304 of 15,160 reviewers) at NeurIPS 2024! (+ Free registration! 😎)
  • 🗞️ [Sep. '24] Two papers got accepted to NeurIPS 2024! 🎉 One is about length generalization of arithmetic Transfomers, and another is about mitigating loss of plasticity in incremental neural net training. See you in Vancouver🇨🇦!
  • 🗞️ [Jun. '24] An early version of our paper on length generalization of Transformers got accepted to the ICML 2024 Workshop on Long-Context Foundation Models!
  • 🗞️ [May. '24] A paper got accepted to ICML 2024 as a spotlight paper (top 3.5% among all submissions)! 🎉 We show global convergence of Alt-GDA (which is strictly faster than Sim-GDA) and propose an enhanced algorithm called Alex-GDA for minimax optimization. See you in Vienna🇦🇹!
  • 🗞️ [Sep. '23] Two papers are accepted to NeurIPS 2023! 🎉 One is about Fair Streaming PCA and another is about enhancing plasticity in RL.
  • 🗞️ [Jan. '23] Our paper about shuffling-based stochastic gradient descent-ascent got accepted to ICLR 2023!
  • 🗞️ [Nov. '22] Our paper about shuffling-based stochastic gradient descent-ascent is accepted to 2022 Korea AI Association + NAVER Autumnal Joint Conference (JKAIA 2022) and selected as the NAVER Outstanding Theory Paper!
  • 🗞️ [Oct. '22] I am happy to announce that our very first preprint is now on arXiv! It is about convergence analysis of shuffling-based stochastic gradient descent-ascent.
  • 🗞️ [Feb. '22] Now I am part of OptiML Lab of KAIST AI.

Education

  • 🏫 Ph.D. in Artificial Intelligence KAIST, Sept. 2023 – Current
  • 🏫 M.Sc. in Artificial Intelligence KAIST, Mar. 2022 – Aug. 2023
  • 🏫 B.Sc. in Mathematical Sciences KAIST, Mar. 2017 – Feb. 2022
    • Minor in Computing Sciences / Summa Cum Laude

Contact & Info

📋 Curriculum Vitae (CV): Here
📧 Email: jhs4015 at kaist dot ac dot kr