Figure 2

Characteristics of encoding models.

(a) Voxel-wise prediction accuracy (in terms of Pearson’s correlation coefficient between measured and predicted brain activity) became saturated with 40–70 encoding channels, and did not increase further with larger numbers of channels. A single line corresponds to the prediction accuracy for one validation set from ten folds in the 10-fold cross-validation procedure from a typical individual (participant 5, bilateral IPG). The results of other individuals and ROIs are presented in Supplementary Figure S1. (b) Distance matrix in terms of the Hamming distance between pairs of code words of the eight view classes (depicted in the left-most inset; white, path; black, wall), under full encoding (‘Full model’), data-driven optimal encoding (‘Optimal model’), data-driven minimum encoding (‘Minimum model’), naïve eight-class encoding (‘Whole-scene model’), and view-part–wise encoding (‘View-part model’). Top panels indicate the Hamming distance normalized by code length and bottom panels are for the original Hamming distance. The histogram beside the full encoding model shows the normalized frequency of the eight scene views in the SC task. Error bars indicate SD over three different maps. Our data-driven models reflected the scene view frequency, a characteristic of map topography, such that more frequently seen views were more frequently predicted. (c) Full encoding and optimal encoding are robust against errors introduced when observing the channels, such that a one bit-inversion error on every encoding channel can be corrected (first row, first and second columns). On the other hand, any single bit-inversion causes unidentifiable codes in whole-scene encoding (fourth column), and leads to misidentification in view-part encoding (fifth column). The second and third rows show the situation when two and three bit-inversion errors occur, respectively. For the other participants’ data, see Supplementary Figure S1.