Canonical Representational Mapping for Cognitive Neuroscience

Understanding neural representations is central to cognitive neuroscience, yet isolating meaningful patterns from noisy or correlated data remains challenging. Canonical Representational Mapping (CRM) is a novel multivariate analysis method to identify neural patterns aligned with specific cognitive hypotheses. CRM maximizes correlations between multivariate datasets - similar to Canonical Correlation Analysis - while controlling for confounding sources of variance, such as shared noise or irrelevant task conditions. We validate CRM with simulations and apply it across diverse neurophysiological datasets: In one application, we use CRM to map large language model activations onto intracranial electroencephalography recordings during story listening, controlling for contextual autocorrelations. In another example, we factorize overlapping representations in functional Magnetic Resonance Imaging into distinct context and episode components. Finally, we uncover frequency-coupled representations shared between hippocampus and medial prefrontal cortex in navigating rodents. Our results introduce CRM as a powerful tool to isolate representations across modalities and species.