Neural Embedding Alignment

Recorded activity patterns can differ because of scale, rotation, or sampling, even when their meaningful relationships are similar. Alignment applies a transformation that reduces these representational differences while retaining structure shared across conditions. This makes comparisons focus on the organization of neural responses rather than measurement-specific coordinate systems, supporting more interpretable analyses across datasets.

Responses to shared stimuli provide corresponding points in the representations being compared. These matched responses give the alignment procedure information about which activity patterns should occupy related positions in a common space. The resulting transformation can then preserve meaningful relationships among those patterns, allowing researchers to evaluate whether different brains, regions, sessions, or models represent the same information similarly.

These approaches provide different ways to learn a transformation between representational spaces. A linear mapping directly relates coordinates, canonical correlation analysis identifies corresponding structure between representations, and orthogonal Procrustes alignment focuses on matching spaces while accounting for geometric differences such as rotation. The choice determines how the comparison handles coordinate changes and which relationships it emphasizes.

A typical workflow begins by collecting responses to common stimuli from the representations that will be compared. Researchers then construct embeddings, select an alignment approach, and learn a transformation between the corresponding activity patterns. After mapping the embeddings into a shared coordinate system, they can compare representational geometry across brains, sessions, regions, experiments, or computational models.

The method is useful when researchers need to compare neural representations across individuals, recording sessions, or brain regions that do not share identical coordinate systems. It also supports integration across experiments and comparison with artificial neural network models. These applications can reveal whether distinct neural systems organize information using similar geometric structures despite differences in recording or model space.

Mapping representations into a shared coordinate system provides a common basis for integrating neural activity from different recording sources and for relating biological data to computational models. Researchers can examine whether activity patterns from distinct modalities or artificial networks preserve similar relationships among stimuli. In neuroscience, this helps test correspondence between biological representation geometry and model-derived structure.