M.S. student, EECS @ GIST · Vision–language models · Mixture of experts · Multimodal AI

I'm a master's student in Prof. Heung-No Lee's INFONET Lab at GIST in Gwangju, South Korea. The thread running through my work is interaction-conditioned computation: learning how inputs (image and text, or pixels and features) should interact, and letting those interactions decide which experts do the work. I apply this mainly to vision–language models, and I've also worked on hyperspectral imaging for face anti-spoofing.

Open to Research Engineer roles in multimodal and vision–language AI — graduating February 2027.

• Vision–language learning and visual reasoning — how image–text interactions shape what a model perceives and reasons about
• Mixture of experts — interaction-conditioned expert weighting, and what the routing actually learns
• Hyperspectral imaging — face anti-spoofing and identity recognition

First author · KICS Winter Conference 2026

A VGG-style CNN with four sparsely gated experts for face anti-spoofing and identification on hyperspectral images (4,374 images, 54 subjects). Raises 4-class live/spoof accuracy from 82.8% to 93.7% (+10.9 pts) over a CNN-only baseline.

Computer Vision course project, GIST 2026

Extends bilinear interaction learning to dense prediction: per-pixel material segmentation on MINC and OpenSurfaces with a ViT-S + ResNet-34 dual encoder and mixture-of-experts modules.

<-- Under review

• J. Kim, Y. Choi, M. Song, S. Kim, S. Lee, H.-N. Lee. IP-MoE: Studying spatial token—patch interactions and dense expert weighting.
• S. Lee, J. Kim, M. Song, S. Kim, H.-N. Lee. Beyond answer preservation: Scientific evidence minimization for private LLM reasoning. -->

Conference and workshop papers

• M. Song, S. Kim, J. Kim, Y. Lee, H.-N. Lee. Retrieval-augmented generation for symbolic folk-art interpretation: Comparing dense and graph-based retrieval using Korean Minhwa. ORACLE Workshop @ EMNLP 2026.
• J. Kim, Y. Choi, H.-N. Lee. An HSI-based CNN-MoE framework for face anti-spoofing and identification. KICS Winter Conference 2026. [code]
• M. Song, S. Kim, J. Kim, H.-N. Lee. AI-assisted dataset creation pipeline for Korean Minhwa. KICS Winter Conference 2026.
• Y. Lee, J. Kim, H.-N. Lee. A prompt-based pipeline for generative game asset synthesis and evaluation. KICS Winter Conference 2026.

• BCAI_A2A — proof-of-concept marketplace where AI agents trade services over the Agent-to-Agent (A2A) protocol, with payments verified on-chain by an EVM smart contract. Course project, GIST 2025, with Minjun Song.
• spectrometer — compressive-sensing spectrometer that reconstructs spectra via non-negative L1-regularized least squares.

• Research: vision–language connector design and training, mixture of experts, parameter-efficient fine-tuning (LoRA), multimodal benchmark evaluation, routing analysis
• Engineering: Python, PyTorch, multi-GPU distributed training (BF16), Docker, Linux

• M.S., Electrical Engineering and Computer Science, GIST, 2025–2027 (expected Feb 2027). Korean Government Scholarship (full tuition and stipend).
• B.A., Political Science and Diplomacy and B.S., Software and Computer Science (double major), Ajou University, 2019–2025.

• 2nd Place, AI Agent & Blockchain Innovation Challenge, The Academic Frontier 2025.

• Volunteer mentor, GIST PIUM (2025–present) — teaching science and offering academic and career mentoring to school students.
• Organizer and speaker, AI Nights, Ajou University's on-campus AI conference (2024) — talk on K-means and DBSCAN.
• Team leader, Daesin-Learning, a university-supported ML/DL study group (2024).