Cross-Subject Neural Alignment via Soft Graph Matching and Dynamic Latent Space Projection for Brain-Computer Interfaces

Individual variability in neural activation patterns presents a major challenge for non-invasive and invasive Brain-Computer Interfaces (BCIs). Direct cross-subject alignmentoften fails because identical sensory stimuli or cognitive tasks can evoke distinct neural rep-resentations across individuals. In this work, we propose a mathematical framework thatleverages topological invariance to align neural responses into a universal latent space Z. Byrelaxing Quadratic Assignment Problems (QAP) to doubly stochastic soft graph matching,incorporating dynamic time-series confidence metrics, and optimizing a joint contrastive-topological loss function, our framework enables universal neural decoding without sacrific-ing individual relational brain structures.