A multi-feature fusion framework combining static lexical representations (W2V) with dynamic contextual representations (GPT) achieved state-of-the-art performance in semantic reconstruction from non-invasive brain recordings. The study benchmarked linear Naive Concatenation against non-linear Multi-Head Cross-Attention, finding a performance hierarchy of Cross-Att > Concat > GPT > W2V. The non-linear cross-attention method demonstrated that neural language decoding
No score is assigned. Sources and their independence are shown in the citation chain below.