Research map / Probabilistic and scientific modeling / Dynamical systems

FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Data assimilation from sparse observations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FAST-Brain, a flow-aligned model that directly predicts clean BOLD signals with a history-conditioned Transformer and a multi-graph GCN spatial prior, and gives a low-dimensional subspace approximation theorem for surrogate rs-fMRI generation.

Problem

Modeling rs-fMRI BOLD dynamics over long horizons while incorporating anatomical spatial structure and exploiting the low-dimensional intrinsic subspace of ambient ROI-time signals.

Approach

Replace flow-matching velocity or noise prediction with direct clean-signal prediction (velocity recovered analytically), combined with a Transformer temporal decoder and a GCN spatial residual branch using SC, fiber length, and functional-network graphs; theory shows the Bayes-optimal denoiser factorizes through the d-dimensional subspace projection.

Key results

• On HCP, FAST-Brain raises FC correlation from 0.726 (best baseline FlowTS) to 0.938 and reduces FC MAE from 0.199 to 0.073.
• On the synthetic RNN dataset, EC recovery reaches correlation 0.956, AUROC 0.999 and AUPRC 0.995, versus 0.851/0.916/0.810 for the best baseline FlowTS.
• Ablations show replacing clean-signal prediction with epsilon- or v-prediction lowers HCP FC correlation to 0.799 and 0.856 respectively.

Surprising findings

• The ablation shows removing the GCN spatial branch barely changes RNN FC (0.932 vs 0.938), while the effect is large on SDDEs and HCP.

Limitations

• Theory covers only the Bayes-optimal denoiser and approximation error; no finite-sample statistical analysis.
• Evaluation on only two synthetic datasets and one real dataset (HCP).
• EC evaluation restricted to synthetic data because ground truth is required.

Experimental setup

Datasets: RNN synthetic dataset SDDEs synthetic dataset HCP rs-fMRI

Benchmarks: RNN synthetic SDDEs synthetic HCP

Metrics: FC correlation FC MAE EC correlation AUROC AUPRC CRPS QICE ProbCorr Conditional FID Discriminative Score

Baselines: NPI DSFM TimeDiff SSSD DiffusionTS FlowTS

Builds on: flow matching EDM NPI DSFM Diffusion-TS

Closest papers by meaning

• Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.87)
• Dynamical Modeling of Behaviorally Relevant Spatiotemporal Patterns in Neural Imaging Data (ICML 2025 · similarity 0.86)
• CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.86)
• BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity (NeurIPS 2025 · similarity 0.86)
• BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.86)
• A scalable generative model for dynamical system reconstruction from neuroimaging data (NeurIPS 2024 · similarity 0.86)