배준현 · Jun-Hyun Bae · Junhyun Bae
AI/ML Postdoctoral Researcher @ Kyungpook National University
Research Interests
My research investigates the robustness and interpretability of deep learning models, with the goal of making their behavior more reliable and controllable. My current focus is mechanistic interpretability, which seeks to characterize how neural networks encode semantic concepts within their internal components. A representative direction is the analysis of cross-attention OV circuits in text-to-image diffusion models, where semantic concepts are represented within low-dimensional spectral subspaces, enabling targeted interventions such as concept removal without retraining. This work builds on my earlier research on robust generalization, including modular architectures for systematic and compositional reasoning and methods grounded in causal inference and invariance for out-of-distribution generalization and debiasing. My broader objective is to connect interpretability with the design of safe and trustworthy models.
Experience
- Kyungpook National University — Postdoctoral Researcher, Department of Artificial Intelligence (Sep 2026 – Present)
- PI: Prof. Heechul Jung
- Carnegie Mellon University — Visiting Scholar (Sep 2022 – Feb 2023) · Full-time, Pittsburgh, PA
- AI-intensive research program, fully funded by the Korean Government (IITP)
Education
- Kyungpook National University — Integrated M.S. & Ph.D. in Artificial Intelligence (Sep 2019 – Aug 2026)
- Dissertation: Mechanistic Decomposition of Cross-Attention OV Circuits in Text-to-Image Diffusion Models
- Advisor: Prof. Heechul Jung | GPA: 4.46/4.5
- Kyungpook National University — B.E. in Electronics Engineering (Mar 2015 – Aug 2019)
- GPA: 4.32/4.5
Publications
- [C5] Jun-Hyun Bae, Wonyong Jo, Jaehyup Lee, and Heechul Jung. “Mechanistic Dissection of Cross-Attention Subspaces in Text-to-Image Diffusion Models.” AAAI Conference on Artificial Intelligence (AAAI), Jan 2026.
- [C4] Jun-Hyun Bae, Minho Lee, and Heechul Jung. “Adaptive Bias Discovery for Learning Debiased Classifier.” Asian Conference on Computer Vision (ACCV), Dec 2024.
- [C3] Jun-Hyun Bae, Chanwoo Kim, and Taeyoung Chang. “Invariant Risk Minimization in Medical Imaging with Modular Data Representation.” International Conference on Electronics, Information, and Communication (ICEIC), Jan 2024.
- [C2] Jun-Hyun Bae*, Taewon Park*, and Minho Lee. “Learning Associative Reasoning Towards Systematicity Using Modular Networks.” International Conference on Neural Information Processing (ICONIP), Nov 2022.
- [C1] Jun-Hyun Bae, Inchul Choi, and Minho Lee. “Meta-Learned Invariant Risk Minimization.” arXiv Preprint, 2021.
Competitions
- NeurIPS 2023 Machine Unlearning Challenge — 8th place (out of 1,121 teams), Google
- AI Hackathon for Fashion Coordination (2020) — 3rd place, ETRI
- AI Hackathon for Speech Recognition (2019) — 10th place, NAVER
Scholarships & Fellowships
- CMU Visiting Scholar Program (2022) — Fully funded by Korean Government (IITP)
- Full-Ride Scholarship (2019 – 2023) — Graduate, Academic Excellence, KNU
- KNU+ Honors Scholarship (2015 – 2019) — Merit-based full-tuition award with semester stipend for outstanding entering students, KNU
Academic Service
Reviewer
- ICML 2026 (Gold Reviewer)
- ICML 2026 Mechanistic Interpretability Workshop
- AAAI 2026
Contact
- Email: junhyun.bae.kr@gmail.com
- GitHub: JunhyunB