2016-nature-Natural speech reveals the semantic maps that tile human cerebral cortex

1. 其他研究文献的缺点

However, all previous studies tested only a handful of stimulus conditions, so no study has yet produced a comprehensive survey of how semantic information is represented across the entire semantic system.

2.本文的方法

Voxel-wise model estimation and validation 

We used a word embedding space to identify semantic features of each word in the stories

The embedding space was constructed by computing the normalized co-occurrence between each word and a set of 985 common English words.   One advantage of voxel-wise modelling over conventional neuroimaging approaches is that the fit models can be validated by predicting BOLD responses to new natural stimuli that were not used during model estimation. This makes it possible to compute effect size by finding the fraction of response variance explained by the models.

Mapping semantic representation across cortex

          We found such a space by applying principal components analysis to the estimated models aggregated across subjects, producing 985 orthogonal semantic dimensions that are ordered by how much variance each explained across the voxels. 

Using PrAGMATiC to construct a semantic atlas

          Given the apparent consistency in the patterns of semantic selectivity across individuals, we sought to create a single atlas that describes the distribution of semantically selective functional areas in human cerebral cortex. To accomplish this, we developed a new Bayesian algorithm, PrAGMATiC, that produces a probabilistic and generative model of areas tiling the cortex.

           The PrAGMATiC algorithm has two components: an arrangement model that determines where functional areas appear on the cortical sheet, and an emission model that determines how the cortical map is produced from an arrangement of areas.

Discussion

One striking aspect of our atlas is that the distribution of semantically selective areas is relatively symmetrical across the two cerebral hemispheres.

Another interesting aspect of these results is that the organization of semantically selective brain areas seems to be highly consistent across individuals.

One limitation of PrAGMATiC as used here is that each area is assumed to be functionally homogeneous.

It should be possible to modify the PrAGMATiC algorithm to model functional gradients explicitly.