Shared Representation Discovery for Multi-Subject Neural Data Analysis

Shared Representation Discovery (ShaReD) is a novel method that identifies neural-behavioral relationships shared across different subjects in multi-subject neural data analysis. This technique is particularly useful when standard cross-subject analyses are limited by the lack of matched time points or anatomical correspondence.

The ShaReD method learns a shared behavioral projection alongside subject-specific neural projections, enabling the identification of common neural representations that are conserved across individuals.

In a recent study, ShaReD was developed and benchmarked using synthetic, primate, and rat data. The method demonstrated its robustness by successfully recovering common structure across various noise levels, sample sizes, and subject counts in the synthetic data. Furthermore, ShaReD was able to separate neural components that were confined to distinct groups of subjects.

When applied to non-human primate motor cortex data, ShaReD successfully identified kinematic representations that generalize across individuals and different reaching tasks with varying movement statistics. Similarly, in rats navigating a spatial alternation task, ShaReD was able to isolate behavior-aligned directions within the CA1-to-PFC communication subspace, highlighting the method's potential to extend multi-subject analysis to datasets with comparable behaviors occurring without cross-subject temporal correspondence.