For decades, cognitive neuroscientists have mapped the human brain by relying on macro-anatomical landmarks—the physical folds, sulci, and gyri of the cortex. Standard group analyses warp every individual's brain template to match a standard anatomical space. However, as Benjamin D. Singer et al. demonstrated in Function-based Intersubject Alignment of Human Cortical Anatomy, this physical alignment blurs fine-grained functional organization. At a granular level, the spatial distribution of functional regions varies significantly across individuals; what represents a specific visual feature or semantic concept in your brain might be physically shifted by millimeters or centimeters in mine.
This inter-individual variability creates a major bottleneck for neural decoding. A machine learning model trained to decode visual stimuli or mental states from one person's brain activity typically fails when tested on another, because their neural representations do not share the same physical coordinates (Alexis Thual et al. in Sample-efficient decoding of visual stimuli from fMRI through inter-individual functional alignment). To overcome this, a powerful paradigm has emerged: functional alignment (or hyperalignment). Instead of forcing brains into a rigid physical template, functional alignment maps neural activity into a shared, high-dimensional representational space based on the information the brain processes.
The Core Mechanism: High-Dimensional Hyperalignment
To understand why functional alignment works, imagine treating each voxel or recording vertex in a specific brain region as a separate dimension in a coordinate system. If you record from 1,000 voxels, you have a 1,000-dimensional space. At any given moment, the brain's pattern of activity can be represented as a single point in this space. As a person experiences a continuous stimulus, their brain activity traces out a trajectory over time.
The foundational insight of hyperalignment, introduced by J Swaroop Guntupalli et al. in A Model of Representational Spaces in Human Cortex, is that when different subjects experience the same complex, naturalistic stimulus—such as watching a movie—their brains trace out trajectories with a shared topology. However, because each subject's voxels are ordered and distributed differently, we view these trajectories through different individual coordinate systems.
Subject A Space (Voxel A1, A2, A3...) Subject B Space (Voxel B1, B2, B3...)
Trajectory X Trajectory X (Rotated/Warped)
\ \
*---> *---> * *---> *---> *
|
Procrustes Alignment
v
Shared Latent Space
Trajectory X
\
*---> *---> *
To align these distinct coordinate systems, researchers use a mathematical tool called the Procrustes transformation. The Procrustes algorithm finds an orthogonal rotation matrix that rotates, scales, and shifts one subject's high-dimensional space to maximize its correlation with another's (Linda Geerligs et al. in A partially nested cortical hierarchy of neural states underlies event segmentation in the human brain).
Crucially, an orthogonal rotation preserves the distances between neural states within a subject. If two movie scenes trigger highly similar neural patterns in Subject A, they will remain highly similar after rotation. By applying these transformations across a group, individual brain geometries are mapped into a common representational space (Chihye Han in High-dimensional structure underlying individual differences in naturalistic visual experience).
Aligning Without a Stimulus: Connectivity-Based Spaces
What if we want to align brains when a shared naturalistic stimulus is unavailable? We can align them using their internal wiring diagrams instead. In fMRI-Based Inter-Subject Cortical Alignment Using Functional Connectivity, Bryan Conroy et al. demonstrated that patterns of functional connectivity—how different regions correlate with one another over time—provide a highly stable basis for registration.
In a connectivity-based alignment framework, a voxel's coordinate is not defined by its physical position, but by its relational profile: a vector of its correlations with the rest of the brain (A computational model of shared fine-scale structure in the human connectome). By mapping these connectivity profiles into a high-dimensional embedding space, Georg Langs et al. showed in Functional Geometry Alignment and Localization of Brain Areas that corresponding functional regions can be identified even when physical anatomy is severely disrupted, such as in patients with brain tumors.
Non-linear and Geometric Extensions
While linear Procrustes rotations are highly effective, they assume a rigid mapping between subjects. The field has evolved to incorporate non-linear transformations and geometric constraints to better capture individual topographies:
- Deep Hyperalignment: To handle complex, non-linear variations, Daoqiang Zhang et al. proposed Deep Hyperalignment, which uses multi-layer neural networks to map heterogeneous neural signals into a unified latent space. This approach captures non-linear functional topographies that linear methods miss, significantly boosting cross-subject decoding accuracy (Shuo Huang et al. in fMRI-based Decoding of Visual Information from Human Brain Activity: A Brief Review).
- Optimal Transport (FUGW): To prevent functional alignment from completely ignoring physical reality, Alexis Thual developed Aligning individual brains with fused unbalanced Gromov Wasserstein (FUGW). FUGW frames alignment as an optimal transport problem that simultaneously minimizes functional differences and anatomical distances, allowing functional regions to vary flexibly in size and shape across individuals while maintaining loose spatial continuity.
Why It Works: The Power of Relational Geometry
Functional alignment works because it shifts the analytical focus from where a signal is located to how signals relate to one another. It uncovers a conserved manifold that captures the universal structure of information processing across distinct physical brains (Conserved structures of neural activity in sensorimotor cortex of freely moving rats allow cross-subject decoding).
By preserving fine-grained spatial details that are typically lost during standard anatomical template-based warping, functional alignment provides several critical advantages:
- Universal Semantic Encoding: It allows researchers to pool data across multiple individuals to build "universal" brain models. In Modeling Semantic Encoding in a Common Neural Representational Space, Cara E. Van Uden et al. showed that semantic encoding models trained in a hyperaligned shared space are more spatially precise and accurate than models trained on a single person's data alone.
- Cross-Subject Stimulus Reconstruction: It enables advanced downstream applications like reconstructing complex visual stimuli or videos from fMRI data across entirely new subjects (Liang Sun et al. in Functional Alignment-Auxiliary Generative Adversarial Network-Based Visual Stimuli Reconstruction via Multi-Subject fMRI; Chong Li et al. in Enhancing Cross-Subject fMRI-to-Video Decoding with Global-Local Functional Alignment).
- Uncovering Fine-Grained Biomarkers: By removing the "noise" of individual anatomical differences, the underlying functional variations become clear. For instance, Ma Feilong et al. demonstrated in The neural basis of intelligence in fine-grained cortical topographies that general intelligence is far better predicted by vertex-level connectivity profiles aligned via hyperalignment than by traditional coarse-grained regional maps. Similarly, Zachary Anderson et al. used hyperalignment transformations in Application of hyperalignment to resting state data in individuals with psychosis reveals systematic changes in functional networks and identifies distinct clinical subgroups to isolate subtle network changes and identify distinct clinical subgroups in psychiatric populations.
Key Foundational Work
Function-based Intersubject Alignment of Human Cortical Anatomy (2009) — contributed 4 concept elements to this direction.
fMRI-Based Inter-Subject Cortical Alignment Using Functional Connectivity (2009) — contributed 1 concept elements to this direction.
A consistent relationship between local white matter architecture and functional specialisation in medial frontal cortex (2006) — contributed 1 concept elements to this direction.