Functional Alignment for Cross-Subject Neural Decoding

This direction is split into other directions: inter-subject functional variability. Its last version stays readable and citable here.

Version 1 of 1120 papers146 conceptsfrozen 2026-10-04

In neuroimaging, the standard way to compare two brains is to treat them like physical objects. We stretch, warp, and rotate one person's brain until its physical landmarks—the folds and ridges of the cortex—line up with a "standard" anatomical template.

However, research by Bertrand Thirion and Gael Varoquaux has shown that this anatomical alignment is fundamentally limited. Even if two brains were physically identical, the way they process information might be spatially shifted. A specific visual concept or linguistic rule might be processed in slightly different physical locations across individuals, a phenomenon known as functional variability.

Functional alignment is the research direction dedicated to moving beyond physical coordinates. Instead of aligning brains based on where they are, it aligns them based on what they do.

The Failure of Anatomical Templates

The core problem is that macro-anatomical landmarks are poor predictors of functional boundaries. As noted in Brain Transfer: Spectral Analysis of Cortical Surfaces and Functional Maps, relying solely on geometric shape often leads to the misalignment of functional circuits.

When we aggregate data from multiple subjects using traditional anatomical warping, we essentially "blur" the signal. If Subject A’s "language center" is 5mm to the left of Subject B’s, a group average will show a weaker, larger, and less precise activation than what actually exists in either individual. This discrepancy is particularly acute in higher-order cognitive regions, such as the language network, where Ben Lipkin and colleagues found that anatomical parcels significantly underestimate effect sizes compared to high-precision functional localization.

Hyperalignment and Shared Representational Spaces

To solve this, researchers developed Hyperalignment. The intuition is to treat brain activity not as a set of 3D voxels, but as a high-dimensional vector space.

In a seminal model of representational spaces, J Swaroop Guntupalli demonstrated that we can use Procrustes transformations—essentially high-dimensional rotations—to map one person's neural response patterns into a "common space" shared by others.

Imagine two people watching the same movie. Their raw voxel activity looks different because their "neural code" is mapped to different physical locations. Hyperalignment finds the transformation that makes their response patterns to the movie as similar as possible. Once in this shared space, a model trained to "read" Subject A's brain can suddenly decode Subject B's thoughts with high accuracy. This has evolved into more complex forms, such as Kernel Hyperalignment, which allows for non-linear relationships between subjects.

Probabilistic and Generative Mapping

While hyperalignment uses rigid rotations, other researchers use probabilistic frameworks to handle the "mismatch" between anatomy and function.

The PrAGMATiC framework, developed by Alexander G. Huth and Thomas L. Griffiths, treats individual brain maps as samples from a shared distribution. It uses a "spring-based" model to allow functional areas to shift, expand, or contract across individuals while maintaining a consistent global topology.

Similarly, the JOSA framework uses deep learning to model the discrepancy between cortical folds and functional activity. By learning this "mismatch" during training, it can achieve state-of-the-art functional alignment at test time using only a standard structural scan, effectively predicting where a person's functional regions should be based on their unique anatomy.

Toward Universal Brain Translators

The ultimate goal of this field is the creation of "universal translators" for neural data. Deep Hyperalignment (DHA), proposed by Muhammad Yousefnezhad and Daoqiang Zhang, uses neural networks to find these mappings, allowing for the decoding of mental states across massive datasets.

Recent work has even moved toward inter-individual neural code conversion without requiring subjects to view the same stimuli. By using deep neural network features as a "common language," researchers can reconstruct images from one person's brain activity using a decoder trained on an entirely different person.

This shift from anatomical correspondence (same location) to informational correspondence (same representational pattern) is what allows modern neuroimaging to pool data across thousands of individuals, significantly boosting the statistical power and generalizability of brain research.

Papers in this version

Showing 60 of 120 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_functional_alignment_for_cross_subject_n_8b66fc55,
  title = {Functional Alignment for Cross-Subject Neural Decoding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/functional-alignment-for-cross-subject-neural-decoding/rdr_8b66fc552fa340c2a07cf067e7d9d788},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_functional_alignment_for_cross_subject_n_8b66fc55_v1,
  title = {Functional Alignment for Cross-Subject Neural Decoding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/functional-alignment-for-cross-subject-neural-decoding/rdr_8b66fc552fa340c2a07cf067e7d9d788/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}