Sangyoon Bae (Stella), PhD Candidate (stellasybae@snu.ac.kr)

Foundation Models • Representation Learning • Scientific ML• Physics-Informed AI • Spatiotemporal Learning• Self-Supervised Learning • Causal Discovery

Education

• [09/2021 ~ ] Doctoral Student in Interdisciplinary Program in Artificial Intelligence, Seoul National University
• [03/2015 ~ 08/2021] BS in Biological Sciences (student designed minor – computational neuroscience for human behavior and cognition), Seoul National University

Research Interests

• Foundation models for scientific spatiotemporal data — embedding the structure of the data-generating process into architecture and pretraining. Methods transfer across domains where measurements are heterogeneous, partially observed, or governed by physical priors: neuroimaging, electrophysiology, climate time-series, and beyond.

Publications

• Bae, S., Azabou, M., Richards, B., Cha, J. (2026). Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining. The Fourteenth International Conference on Learning Representations (ICLR 2026) [link]
• Bae, S., Kwon, J., Yoo, S., Cha, J. (2026). Learning brain dynamics across distinct scaling regimes reveals psychiatric signatures. Communications Biology [link]
• [Under Review, NeurIPS 2026] Bae, S., et al. PIMSM: Physics-Informed Multi-Scale Mamba for Stable Neural Representations under Distribution Shift. [arXiv]
• [Under Review, NeurIPS 2026] Bae, S., Oprescu, M., Park, D., Yoo, S., Cha, J. (2026). Latent-Space Causal Discovery from Indirect Neuroimaging Observations. [arXiv]
• Watkins, W., Wang, H., Bae, S., et al. (2024). Quantum Privacy Aggregation of Teacher Ensembles (QPATE) for Privacy-preserving Quantum Machine Learning. ICASSP 2024 [link]
• Kim, P., Kwon, J., Joo, S., Bae, S., et al. (2024). SwiFT: Swin 4D fMRI Transformer. Advances in Neural Information Processing Systems, 36 [link]

Awards

• Best Paper Award, IEEE QCE (2025)
• IPAI Development Fund Scholarship, SNU (2024)
• Sung-Am Educational Foundation Scholarship (2023)
• Graduate School of AI Fellowship, SNU (2021–2023)

“Structure Inductive Bias

is Necessary Condition

for Scientific FM”

MBBN : Learning brain dynamics across distinct scaling regimes reveals psychiatric signatures | Communications Biology, 2026

[Model Procedure]

[Pretraining Model Architecture]

• Encoded BOLD signals into three frequency bands reflecting distinct fractal properties.
• Designed communicability-based pretraining loss for transfer learning on large-scale datasets.

[Results]

[Interpretation]

• Built self-attention-weighted connectivity and GradCAM-style interpretation to detect frequency-specific biomarkers.

• MBBN outperforms most models in various tasks.
• Model trained with communicability-based masking loss outperforms in transfer learning.

MBBN : Learning brain dynamics across distinct scaling regimes reveals psychiatric signatures | Communications Biology, 2026

POYO-CAP : Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining | ICLR 2026

[Framework]

[Results]

• 12–13% relative improvement over from-scratch training on Allen Brain Observatory.

[Results (cont’d)]

• Enables smooth, monotonic scaling with model size (baselines plateau/destabilize).

POYO-CAP : Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining | ICLR 2026

• High-fidelity movie frame reconstruction from calcium recordings without external stimulus information.

• 1.98× data efficiency gain vs. mixed-population training.

• SST, VIP, PVALB, NTSR1 lines identified as predictable (near-Gaussian, low kurtosis).

Q-PATE : Quantum Privacy Aggregation of Teacher Ensembles | ICASSP 2024

SwiFT : Swin 4D fMRI Transformer | NeurIPS 2023

PIMSM : Physics-Informed Multi-Scale Mamba for Stable Neural Representations under Distribution Shift | Under Review, NeurIPS 2026

[Architecture and Results]

[Key Contributions]

• Temporal Kernel Mismatch Problem

Single-scale SSMs become structurally mismatched under acquisition shifts (TR, sampling rate), inducing representation drift and degraded transfer.

• Physics-Informed Temporal Hierarchy

SpectralHyperNet estimates knee frequencies from PSD; anchors Mamba Δ to acquisition time units. Multi-scale temporal kernels grounded in neural scaling laws (piecewise power-law).

• 4-Axis Robustness Evaluation

L1 TR shift (HCP), L2 brain-state (rest→task fMRI), L3 sitcom→documentary, L4 cross-domain WEATHER-5K. Monotonic scaling; robust under extreme low-data (~1% supervision).

INCAMA : Latent-Space Causal Discovery from Indirect Neuroimaging Observations | Under Review, NeurIPS 2026

[Framework and Results]

[Key Contributions]

• Physics-Aware Inversion

Separates latent neural activity from measurement artifacts: HRF deconvolution for fMRI; leadfield inversion for EEG. Modality-specific priors with formal inversion recovery guarantees.

• Latent-Space Causal Discovery

Mamba-based temporal encoding for delay-aware causal inference. Identifies sparse directed graphs (nonstationarity as soft interventions) with latent-space identifiability guarantees.

• Zero-Shot Transfer to HCP

Without fine-tuning, recovers canonical visuo-motor pathways (V1→V2, M1↔S1) consistent with neuroanatomy. Outperforms standard pipelines on large-scale biophysical simulations.