Sangyoon Bae

Sangyoon Bae

Scientific ML · Physics-Informed Architectures · Foundation Models for Dynamical Systems
Ph.D. Candidate · Interdisciplinary Program in Artificial Intelligence
Seoul National University
Email GitHub LinkedIn Scholar

I build physics-informed foundation models for scientific systems — embedding the structure of the data-generating process directly into architecture and training. My focus is scientific time-series: signals acquired indirectly and often unstructured, where physics-grounded inductive biases are what turn latent structure into a learning signal.

2 sole first-author papers (ICLR 2026, Communications Biology) · 2 sole first-author papers under review · 1 co-first-author paper (ICASSP 2024) · 7 peer-reviewed publications · Models evaluated on datasets up to 49k+ subjects across neural, climate, and population-scale data

News
Publications
🔬
ICLR 20262026

Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining

Bae, S., Azabou, M., Richards, B., Cha, J.
paper
🧠
Communications Biology (Early Access)2026

Learning brain dynamics across distinct scaling regimes reveals psychiatric signatures

Bae, S., Kwon, J., Cha, J., Yoo, S.
paper
📈
arXiv Preprint2026

PIMSM: Physics-Informed Multi-Scale Mamba for Stable Neural Representations under Distribution Shift

Bae, S., Yoo, S., Cha, J.
arXiv Preprint
🔗
arXiv Preprint2026

Latent-Space Causal Discovery from Indirect Neuroimaging Observations

Bae, S., Oprescu, M., Park, D., Yoo, S., Cha, J.
arXiv Preprint
⚛
IEEE QCE 20252025

Resting-state fMRI Analysis using Quantum Time-series Transformer

Park, J., Bae, S., Seo, J., Chen, S., Tseng, H., Cha, J. & Yoo, S.
paper
🌊
NeurIPS 20232023

SwiFT: Swin 4D fMRI Transformer

Kim, P., Kwon, J., Joo, S., Bae, S., Lee, D., Jung, Y., ... & Moon, T.
paper
🔐
ICASSP 20242024

Quantum Privacy Aggregation of Teacher Ensembles (QPATE) for Privacy-preserving Quantum Machine Learning

Watkins, W., Wang, H., Bae, S., Tseng, H. H., Cha, J., Chen, S. Y. C., & Yoo, S.
paper
🖼
arXiv Preprint2024

Macro2Micro: Cross-modal Magnetic Resonance Imaging Synthesis Leveraging Multi-scale Brain Structures

Kim, S., Kwon, J., Kwon, J., Bae, S., Lin, Y., Yoo, S., & Cha, J.
arXiv Preprint
📡
IBRO Neuroscience Reports 20232023

Resting-state theta phase-gamma amplitude coupling as a biomarker for the transdiagnostic dimensional approach in psychiatric disorders

Kim, M., Shim, Y., Kwon, J., Bae, S., … & Kwon, J. S.
paper
Selected Research
PIMSM

Physics-Informed Multi-Scale Mamba for Dynamical Systems

  • Formalizes temporal kernel mismatch as a failure mode under distribution shift; derives scale-specific SSM discretization from spectrum-estimated knee frequencies with A-scale regularization
  • HCP fMRI: highest representation stability (CKA 0.812) under temporal truncation; strongest low-resource accuracy at 1% supervision
  • Weather-5K: lowest variable-wise MAE across all horizons on held-out-station OOD forecasting (15.2% avg MAE reduction vs. best baseline)
CKA 0.812 on HCP · 15.2% MAE reduction on Weather-5K · ~1% labeled data
POYO-CAP

Curriculum Pretraining for Heterogeneous Scientific Data

  • Identified “scaling collapse” in SSL pretraining on heterogeneous scientific data: indiscriminate training degrades performance as noisy signals dominate the loss
  • Solution: use higher-order statistics (skewness/kurtosis) to identify statistically regular subsets; domain-informed curriculum — pretrain on predictable signals first, fine-tune on stochastic ones
  • Unlocks monotonic scaling and 1.98× data efficiency; validated on 200k+ calcium imaging recordings
SSIM +13% · 1.98× data efficiency · 200k+ sensor recordings · 4×V100
INCAMA

Physics-Aware Causal Discovery from Indirect Measurements

  • Recasts causal discovery from indirect neuroimaging as a conditional identifiability problem: physics-aware inversion (HRF deconvolution, EEG source localization) coupled with delay-aware Mamba-based latent causal discovery
  • Proved identifiability reduction with inversion-error propagation bound
  • 2–3× F1 improvement over baselines on TVB simulations; zero-shot transfer to HCP motor-task fMRI recovers canonical visuo-motor pathways with 4× higher recall
>3× improvement vs. DL baselines · zero-shot transfer (arXiv:2602.09034)
MBBN

Physics-Derived Multi-Band Attention for fMRI

  • Derives attention mechanism from the physics of scale-free power-law spectra: decomposes fMRI signals into domain-informed frequency bands, applies specialized self-attention per band
  • Physics-grounded pretraining loss (network communicability) anchors optimization
  • Trained at scale on 49,673 individuals; up to 41.36% AUROC improvement on psychiatric classification
AUROC +41% on ASD classification · 49k+ subjects (Communications Biology)
Work Experience
2024

MILA — Montréal Institute of Learning Algorithms

Visiting Researcher · Neuro-AI Foundation Models & Biologically-Grounded Modeling
Advised by Prof. Blake Richards
Collaborated on scalable foundation models for scientific data. Developed the data-centric curriculum insight — statistical regularity as a pretraining criterion — that became the core contribution of POYO-CAP (ICLR 2026).
2023

Brookhaven National Laboratory

Visiting Researcher · Quantum Machine Learning Infrastructure
Advised by Dr. Shinjae Yoo
Applied privacy-preserving techniques to federated quantum learning settings; resulted in ICASSP 2024: “Quantum Privacy Aggregation of Teacher Ensembles (QPATE).”
2022

Institut de Neurosciences des Systèmes, Aix-Marseille Université

Visiting Researcher · Adaptation of The Virtual Brain (TVB) to ADHD Progression
Advised by Prof. Viktor Jirsa
Worked on simulator-based inverse modeling with The Virtual Brain (TVB); experience directly seeded the INCAMA project on physics-aware causal discovery.
Education

Ph.D. Candidate — Interdisciplinary Program in Artificial Intelligence

Seoul National University, Seoul, South Korea  ·  September 2021 – Present

B.S. — Biological Sciences

Student-Designed Minor: Computational Neuroscience for Human Behavior and Cognition
Seoul National University, Seoul, South Korea  ·  March 2015 – August 2021

Technical Skills
Machine Learning
PyTorch (expert), JAX/Flax (XLA compilation & JIT optimization), distributed & multi-GPU training, large-scale pretraining, foundation model development (SSL, curriculum learning, scalable architectures)
Scientific ML
Physics-informed neural networks, inverse problem theory, physics-aware inductive biases, state-space models (Mamba/SSM), multifractal signal analysis, biophysical forward models, causal discovery, dynamical systems
Data Modalities
fMRI, EEG, calcium imaging, large-scale population datasets (49k+ subjects, 200k+ neurons)
Systems
Multi-GPU training (4× V100), HPC clusters, CUDA; biophysical simulators (TVB)
Scholarships & Awards
Mentoring

Mentored 8 undergraduates, master’s students, and high-schoolers across neuroscience and AI. Led weekly reading groups and code reviews; resulted in peer-reviewed publications.