Each volume consisted of 15 ∗ 6 mm thick slices with

an inter-slice gap of 1 mm; FOV: 20 ∗ 20 cm; size of acquisition matrix, 64 ∗ 64; NEX: 1.00. The parameter values of the anatomical scans were TR = 7.284 ms, TE = 2.892 ms, FA = 11 degrees, bandwidth = 31.25 kHz, and voxel size = 1 mm isotropic. Following the settings used by Mitchell et al., we used oblique slices in the sagittal view with a tilt of −20 to −30 degrees such that the most inferior slice was above the eyes (anteriorly) and passed through the cerebellum (posteriorly). The fMRI pre-processing was performed with SPM8 (Welcome Department of Imaging Neuroscience, UK). Corrected for motion was applied to the images, followed by co-registration of functional and anatomical images, segmentation to identify grey matter, and normalisation into standard Montreal signaling pathway Neurological Institute (MNI) BI 2536 order spaces at a re-sliced voxel size of 3 × 3 × 6 mm. The unsmoothed data were analysed with the Searchlight method. The computation for the Searchlight was made using PyMVPA2.0,

a Python package intended to run machine-learning programs applied to human neurological data. Searchlight yielded an accuracy map for classification of the stimulus language in each trial (Korean or Chinese script) with the voxels with higher accuracy indicating small local regions that are more informative. In our study, the method was applied to the entire brain, over spherical regions of radius 3. The machine-learning classifier from used with Searchlight was a logistic regression with L2-norm regularisation (also termed ridge regression or Tikhonov regularisation). Consecutively, the z-statistic of the accuracy for each voxel was computed and screened out with a threshold of 3.08, corresponding to a p-value of 0.001 under the hypothesis of normal distribution. Participant-based images were visualised using the xjView toolbox (http://www.alivelearn.net/xjview) to produce sensitivity maps analogous to statistical maps

of a GLM. xjView toolbox was also used for extracting clusters of informative voxels in which the discrimination accuracy was high. For the GLM analysis (Friston et al., 1994 and Friston et al., 1995) the data were additionally smoothed using an 8 mm Gaussian kernel. A conventional General Linear Model contrastive analysis was performed for each individual participant. The group-averaged effects were computed using a fixed-effects model. For the group analysis, those clusters of 4 or more that were above a threshold of p < 0.05 FWE (at both the cluster-level and peak-level) were considered to be significant. This work was supported by a grant from Kaken, Japan Society for the Promotion of Science (JSPS), Kiban (C)-23500171.