spectral-ot-functional-alignment

name

spectral-ot-functional-alignment

description

SpectralOT — a geometry-aware, spectral optimal-transport functional alignment method for fMRI that embeds cortical surface geometry (Laplace-Beltrami eigenmodes) into the alignment cost to regularize cross-subject alignment while preserving anatomical structure. Use when building population-level brain decoders, doing cross-subject fMRI alignment, or need a fast geometry-preserving alternative to Hyperalignment / Riemannian alignment. Trigger words: functional alignment, cross-subject decoding, fMRI alignment, Laplace-Beltrami, cortical geometry, optimal transport, Hyperalignment, surface-based alignment, population decoder.

SpectralOT — Geometry-Aware Functional Alignment for fMRI

What it is

SpectralOT (arXiv:2607.10931, Barbarant et al., 2026-07-12) is a new functional alignment method for fMRI that aligns functional data across individuals before training population-level decoders. The core innovation: it embeds cortical surface geometry into Laplace–Beltrami eigenmodes and uses them to regularize an optimal-transport (OT) alignment, striking the balance between (a) aligning functional features and (b) preserving anatomical structure, while staying computationally efficient (spectral/spectral-OT rather than full pairwise OT on vertices).

Problem it solves

Inter-individual variability in brain response patterns limits decoders that generalize across subjects. Functional alignment maps each subject's data into a shared space. Existing methods trade off alignment quality vs. anatomical preservation vs. speed:

• Hyperalignment (Haxby 2011): fast, PCA-based, but ignores cortical geometry.
• Riemannian / surface-based alignment: preserves geometry but is slow / optimization-heavy.
• Optimal transport on vertices: geometrically natural but expensive at brain scale.

SpectralOT addresses all three: spectral (Laplace–Beltrami) regularization + OT formulation + efficient solver.

Core methodology (reusable recipe)

1. Build cortical geometry basis. Compute the Laplace–Beltrami (LB) eigenmodes of each subject's cortical surface mesh (e.g. from FreeSurfer / fsLR surface). These eigenmodes are the "spectral" coordinates that encode anatomy independent of functional data.
2. Functional data projection. Represent each subject's fMRI time series / feature maps in a shared functional basis (e.g. task contrasts or ROI responses).
3. Regularized optimal transport. Solve an OT problem that transports one subject's functional embedding toward another's, with a geometry penalty term measured in LB-eigenmode space (penalize warps that violate cortical adjacency/structure). This is the "SpectralOT" cost: cost = functional_OT_distance + λ · geometry_deviation(LB_modes).
4. Shared decoder training. After aligning all subjects into the common space, train a single population-level decoder (e.g. ridge / logistic) on the pooled aligned data.
5. Leave-one-subject-out evaluation. Align held-out subject to the rest; decode; report cross-subject generalization vs. baseline (no alignment / Hyperalignment).

When to use

• Multi-subject fMRI studies needing a population decoder (decoding thoughts/stimuli across people).
• Scenarios where anatomical preservation matters (surface-based ROIs, cortical topography).
• Compute-constrained pipelines (need speed but want geometry awareness).
• Benchmarking new alignment methods against a strong, simple, geometry-regularized baseline.

When NOT to use

• Single-subject analyses (alignment is only meaningful across subjects).
• Non-surface data (volume-only fMRI without a mesh) — you'd need to adapt the LB basis.
• If a heavyweight learned alignment (deep hyperalignment, VPN) is already beating spectral methods on your task and compute is not a concern.

Implementation notes / pitfalls

• Mesh consistency is mandatory. All subjects must be in the same surface template (fsLR32k / fsaverage) so LB eigenmodes are comparable. Mis-registered meshes silently corrupt the geometry term.
• λ tuning. The geometry-weight λ controls the alignment↔anatomy trade-off. Sweep λ and validate on a held-out decoding metric, not on alignment loss alone.
• LB eigenmode count. Use enough low-frequency modes to capture coarse anatomy but not so many that you overfit to noise; 100–300 modes is a typical starting range.
• OT solver. Use an entropic-regularized Sinkhorn OT for speed; the geometry penalty can be folded into the cost matrix before Sinkhorn.
• Functional basis choice strongly affects results — task-contrast vectors are the standard for Natural Scenes / Haxby-style datasets; for resting-state use parcellated FC vectors.

Relationship to existing ai_collection skills

• Distinct from brain-alignment-* (those are DNN↔brain representational alignment, RSA-style). SpectralOT is subject↔subject functional alignment for decoding, not model↔brain.
• Complements atlas-free-brain-network-transformer (single-subject spatial) — SpectralOT is the cross-subject bridge.
• Overlaps thematically with flexibrain-resolution-agnostic-fmri-encoding (encoding side) — pair them: align with SpectralOT, then train encoding models.

Verification

• Reproduce the paper's cross-subject decoding gain on a public dataset (Haxby, Natural Scenes Dataset, or fMRI Nimstim).
• Ablate the geometry term (λ=0) → should degrade to near-Hyperalignment behavior; confirms the LB regularization is doing work.
• Check that aligned maps preserve known retinotopic / functional gradients (sanity anatomical check).

Source

Barbarant P-L, Meyniel F, Thirion B. "Fast Whole-Brain, Geometry-Aware Functional Alignment for Cross-Subject Decoding." arXiv:2607.10931 (2026-07-12), q-bio.NC.