Fig. 6.

Semantic organization in cortex according to Manning et al. [35]. (a) Mean power at specific frequencies are calculated from neural recordings during word presentation and recall and then passed through principal components analysis (PCA) [35]. The PCA components that vary systematically with the meanings of the presented words are used to construct a neural similarity matrix by measuring the pair-wise cosine similarities between the PCA feature vectors for each word [35]. (b) Semantic features of the words themselves are calculated through latent semantic analysis (LSA) [35]. A semantic similarity matrix is constructed by measuring the pair-wise cosine similarities between the LSA feature vectors for each word [35]. Note that the neural similarity (a) displays a distinct band-diagonal structure, implying near-optimal self-similarity [35]. This, along with the finding that neural and semantic similarity are highly correlated just prior to word recall, suggests that words are encoded in cortex based on semantic features during speech intention [35]. Figure reused with permission from Ref. [35]