Cross-Subject Semantic Decoding via Shared-Space Alignment

Generalizing insights from invasive neural recordings across different individuals is a significant challenge due to the inherent variability in electrode placements, anatomical structures, and neural signal patterns. This study introduces a novel cross-subject semantic decoding framework designed to overcome these inter-subject differences. The core idea involves aligning neural responses, specifically those related to speech perception, from multiple subjects into a common, shared latent space. Within this shared space, a decoder is trained to map the aligned neural representations to contextual semantic embeddings. For a new, unseen subject, the framework estimates a subject-specific projection into this predefined shared space, allowing the pretrained decoder to be applied directly without any further retraining. Experimental results, using electrocorticography data collected during natural language comprehension, consistently show that this approach outperforms existing baseline methods. It significantly reduces the performance drop when testing on held-out subjects, demonstrating improved cross-subject generalization by effectively capturing shared stimulus-related representations while mitigating individual-specific neural variations.