New Framework Improves Cross-Subject Neural Semantic Decoding

A new framework addresses the challenge of generalizing neural decoding across different individuals, a significant hurdle in invasive neural recording research due to variations in brain anatomy and signal patterns. The proposed method, called cross-subject semantic decoding, aligns neural responses related to speech perception from multiple subjects into a common latent space. This alignment is achieved using a shared response model, after which a decoder is trained to predict contextual semantic embeddings from these projected neural responses. For new, unseen subjects, the framework estimates a subject-specific projection into the shared space, allowing the pretrained decoder to be applied directly without further training. Experimental results using electrocorticography data demonstrate that this approach consistently outperforms existing methods, significantly reducing performance drops when generalizing from source subjects to new individuals. This suggests that aligning neural activity into a shared semantic space is an effective strategy for overcoming inter-subject variability and improving generalization in neural representation learning.