Aligning Cortical Activity via Shared Functional Spaces

49 concepts47 papers1 proposal

Traditional neuroimaging relies on a foundational assumption: if we align two brains geometrically based on their physical anatomy, we can compare their functional activity voxel by voxel. To achieve this, researchers typically normalize individual brain scans to a standard structural template using macro-anatomical landmarks like cortical folds, sulci, and gyri.

However, this assumption breaks down at finer spatial scales. As shown by Lombaert et al. and Thirion et al., physical topography does not consistently map to functional boundaries across different individuals. Even with precise structural registration, the exact spatial coordinates of functional activation can vary significantly between subjects (Dohmatob et al.; Xu et al.). This phenomenon—known as anatomical-functional misalignment—means that fine-grained neural patterns are frequently discarded as idiosyncratic noise when aggregating data across a group (Guntupalli et al.), severely limiting the performance of cross-subject neural decoding models (Huang et al.).

To resolve this, a powerful paradigm has emerged: instead of forcing diverse brains into a rigid, shared physical coordinate system, we can project their functional activity into a high-dimensional shared representational space.


The Intuition: From Physical Space to Representational Space

To understand why functional alignment works, imagine two people watching the same movie while their brain activity is recorded via fMRI.

If we look at a specific region of interest containing $N$ voxels, we can represent the brain's state at any given second as a single point (or vector) in an $N$-dimensional space, where each axis corresponds to the activation level of one voxel. As the movie plays, the changing neural activity traces a trajectory—a cloud of points—through this high-dimensional space.

Subject A's Voxel Space              Subject B's Voxel Space
       Voxel 2                              Voxel 2
         ^                                    ^
         |   * (Scene 2)                      |       * (Scene 1)
         |  /                                 |      /
         | /                                  |     / 
         |/_____* (Scene 1)                   |____/_____* (Scene 2)
         +----------> Voxel 1                 +----------> Voxel 1

Because Subject A and Subject B have different physical voxel layouts and functional boundaries, their point clouds look completely different when plotted on their respective physical voxel axes. A machine learning model trained to decode what Subject A is seeing based on Voxel 1 and Voxel 2 will fail completely if applied directly to Subject B's voxels.

However, the geometric relationships between the points within each cloud are remarkably preserved. For both subjects, the neural vector for a "happy scene" might be close to the vector for an "excited scene," but far from the vector for a "scary scene."

Functional alignment algorithms exploit this shared geometry. They treat the problem as a high-dimensional manifold alignment task, finding a transformation matrix (such as a rotation) that aligns Subject B's point cloud directly onto Subject A's point cloud, or onto a common template space. Once aligned in this abstract representational space, the physical coordinates no longer matter; the underlying informational profiles match perfectly.


Mathematical Frameworks for Functional Mapping

Linear Hyperalignment

The foundational implementation of this concept is searchlight hyperalignment, introduced by Guntupalli et al.. This approach uses the classical Orthogonal Procrustes problem to calculate a rotation matrix $R_i$ for each subject $i$ that minimizes the distance between their response matrix $X_i$ and a common template matrix $M$:

$$\min_{R_i} \| X_i R_i - M \|_F \quad \text{subject to} \quad R_i^T R_i = I$$

By constraining $R_i$ to be an orthogonal matrix, the transformation preserves the geometry of the representational space (i.e., it rotates the point cloud without distorting the distances between vectors). The authors demonstrated that this common space allows researchers to build universal semantic encoding models: a model trained on data from a pool of subjects can accurately predict how a completely new individual's brain will respond to complex concepts (Van Uden et al.).

Non-linear and Deep Extensions

While linear rotations preserve geometry, they assume that the underlying functional variations across subjects can be resolved through rigid transformations. To capture more complex, non-linear functional discrepancies, Zhang et al. proposed Deep Hyperalignment (DHA). DHA utilizes deep neural networks to learn non-linear mapping functions that project individual subject spaces into a shared space, significantly improving decoding accuracy for highly complex mental states. To handle massive datasets with many subjects, Xu et al. introduced Gradient Hyperalignment, which optimizes the alignment mapping via gradient descent, making the process highly scalable.

Supervised Hyperalignment

Standard hyperalignment is unsupervised, relying entirely on the shared temporal trajectory of a continuous stimulus (like a movie) to align the spaces. However, when explicit task labels are available, incorporating them can guide the alignment toward features that matter most for cognitive state classification. To leverage this, Zhang developed Supervised Hyperalignment (SHA). By integrating task labels directly into the optimization objective, SHA ensures that the resulting shared space maximizes the separability of different cognitive classes, boosting downstream multi-subject decoding performance.


Beyond Human Subjects: Evolutionary Alignment

The power of functional alignment is not restricted to mapping human brains to other human brains. Because it relies on the geometry of informational profiles rather than physical landmarks, it can bridge entirely different species.

Xu et al. utilized a joint-embedding functional alignment method to map human brains to macaque brains based on functional connectivity profiles rather than cortical anatomy. This cross-species functional alignment revealed an evolutionary hierarchy within the connectome: while primary sensory systems remain highly conserved and structurally aligned across primates, high-level associative networks—such as the Default Mode Network—have functionally drifted away from their ancestral anatomical positions during human evolutionary expansion.

Key Foundational Work

Analyzing for information, not activation, to exploit high-resolution fMRI (2007) — contributed 2 concept elements to this direction.
Exploring Functional Connectivities of the Human Brain using Multivariate Information Analysis (2009) — contributed 1 concept elements to this direction.
Detection of Spatial Activation Patterns as Unsupervised Segmentation of fMRI Data (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_inter_subject_functional_variability_1990fa00,
  title = {inter-subject functional variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-subject-functional-variability/rdr_1990fa004c0c4033ae70f56e102ef000},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_inter_subject_functional_variability_1990fa00_v1,
  title = {inter-subject functional variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-subject-functional-variability/rdr_1990fa004c0c4033ae70f56e102ef000/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake